Gemma 3n #39059
Merged
Gemma 3n #39059
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
* initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com>
4 tasks
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
zaristei
pushed a commit
to zaristei/transformers
that referenced
this pull request
Sep 9, 2025
* Gemma 3n * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix import test * add model args * auto_docstring * replace test path * consistency * skip tests for now * fix docstring for doc builder * skip unused attr --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>
zaristei
pushed a commit
to zaristei/transformers
that referenced
this pull request
Sep 9, 2025
* Gemma 3n * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix import test * add model args * auto_docstring * replace test path * consistency * skip tests for now * fix docstring for doc builder * skip unused attr --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>
zaristei
pushed a commit
to zaristei/transformers
that referenced
this pull request
Sep 9, 2025
* Gemma 3n * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix import test * add model args * auto_docstring * replace test path * consistency * skip tests for now * fix docstring for doc builder * skip unused attr --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>
zaristei
pushed a commit
to zaristei/transformers
that referenced
this pull request
Sep 9, 2025
* Gemma 3n * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix import test * add model args * auto_docstring * replace test path * consistency * skip tests for now * fix docstring for doc builder * skip unused attr --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>
zaristei
pushed a commit
to zaristei/transformers
that referenced
this pull request
Sep 9, 2025
* Gemma 3n * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix import test * add model args * auto_docstring * replace test path * consistency * skip tests for now * fix docstring for doc builder * skip unused attr --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>
zaristei
pushed a commit
to zaristei/transformers
that referenced
this pull request
Sep 9, 2025
* Gemma 3n * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix import test * add model args * auto_docstring * replace test path * consistency * skip tests for now * fix docstring for doc builder * skip unused attr --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>
zaristei
pushed a commit
to zaristei/transformers
that referenced
this pull request
Sep 9, 2025
* Gemma 3n * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3p5RMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3p5 overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3p5 text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * regenerating modeling file after syncing to HEAD * Use torch.std(..., unbiased=False) for activation sparsity (huggingface#8) * Refactoring to a single QVK Norm (huggingface#13) * AltUp: support scale_corrected_output (huggingface#14) * Converts einsums to nn.Linear (huggingface#7) * Converts einsums to nn.Linear * Removing unused variables * Aligning SharedKVCache with HybridCache (huggingface#11) * Alinging SharedKVStore with HybridCache * Remove KVStore. Refactor apply_rotary_pos_emb for sharing * Addressing review comments * Supporting split modality embeddings in Gemma3n (huggingface#10) * Adding the Embedder class * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation * Apply suggestions from code review Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Update modular Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> * Addressing review comments, prop drilling audio and vision configs to the text config * Removing TODO's that have been addressed * Simplify Embedder init and add audio embeddings * Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder * Refactoring vision and audio embeddings into ConditionalGeneration model --------- Co-authored-by: Ryan Mullins <ryan@ryanmullins.org> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating attention mask for Gemma 3.5 (huggingface#15) * xxx_token_index to xxx_token_id * remvoing deprecated last_cache_position * Removing references to SigLIP * Always init per-layer inputs * Using torch.finfo().min for epsilon_tensor * Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas * fix modular GEMMA3N_INPUTS_DOCSTRING * Gemma3nAttention inherits from Gemma3Attention * Modular inheritance fixes * CausalLM conversion script for 4B model (huggingface#16) * Add Gemma3n Audio Encoder (huggingface#6) * initial commit of Gemma 3.5 scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding gemma3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3.5 (huggingface#3) * Initial Gemm3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3.5 * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right Gemma 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3.5 * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * CausalLM conversion script for 4B model * inv_timescales to non-persistent buffer * Addressing audio encoder Attention feedback * Addressing Gemma3nAudioSSCPConvBlock feedback * Addressing Gemma3nAudioConformerAttention feedback * Addressing padding feedback * Weights conversion loads audio state dict * Always use vision_config so saving works * Token id updates for configs * Stubs for interleaving audio embs * Addressing reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Fixing cache access error * Removing duplicate code from a bad merge * Gemma 3n Text + Vision Part 1 (huggingface#17) * testing utilities for numerics comparisons * Corrected einsum to nn.Linear weights conversion * Inherit scaled word embs from Gemma3 not Bart * Fixing transposes for collapsed linears * More transpose fixes * numpy api fix * RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True * Force AltUp to float32 * Updating debugging script for AudioEncoder debugging * Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs * Correcting attention einsum conversions * RMSNorm in type of x * Fixing douplicate laurel norm/gating * KV sharing using the right previous indices * Refactor kv shared index computation. Correct frac_shared_layers * Use num_shared_layers instead of inferring from a fraction * fixing a bug for logging * Fix shared data_ptrs in altup inits * rope: adjust proj -> norm -> rope to preserve computation (huggingface#20) * rope: adjust proj -> norm -> rope to preserve computation * Removing some breaking language model fluff in ConditionalGeneration * Consolidate query_states transforms --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Vectorize the loops in AltUp (huggingface#19) * Vectorize the loops in AltUp * fix typo * Expanding to support batched inputs * remove extra debug script * Fix AltUp.forward --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel * Convert norm to 1/sqrt (huggingface#21) * Convert norm to 1/sqrt * Scale shift change per Phil's rec * Adding default activation sparsity * Fixing 2B config in weights conversion script * Fixing RMSNorm parameters - adding scale_shift and with_scale * Correcting query pre-attention scaling * Adding query_rescale_scalar to text config * Adding layer_idx to MLP * Permafix for input_layernorm * Use 1/sqrt instead of rsqrt in DecoderLayer * Fix o_proj conversion * Conversion script update for vision encoder * Removing logging for debugging timm model * Fixing bugs in Gemma3nForConditionalGeneration for text generation * Generating the modeling_gemma3n.py file * Removing the addition of an erroneous line in the modeling file * Adding gemma3n text model to modeling_auto * Bugfix: Updating the interleaving of inputs_embeds and vision_embeds * Updating the modeling file with the latest bugfix changes * Updating models/auto for Gemma 3n * using AutoTokenizer in forward test * Adding processing_gemma3n.py * Gemma 3n configured for AutoModel. Conversion script updated. * Removing errant merge artifacts --------- Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> * Removing errant debugging statements from Gemma 3 * Gemma3n audio model (huggingface#18) * testing utilities for numerics comparisons * Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock * Add audio version of forward script based on RyanMullins' implementation * Updating to match encoder tests. WIP: config question needs resolving * Updates to audio classes to enable end-to-end running * Removing vestigial classes, cleaning up print statements * Adding SiLU / Swish to audio conformer feed forward block * Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio * Adding outputs to audio test * Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model * Update forward test to load from local weights * Update conversion to process / output audio layers * Update __all__ to export audio encoder * AutoModel registration for Gemma 3n Audio * Use AutoModel for ConditionalGeneration.audio_tower * Fixing input_proj_linear transpose * Fixing Gemma3NanoAudioConformerAttention.post conversion * Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion * Correcting indentation issue on Gemma3p5RMSNorm --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Text + Vision Part 2 (huggingface#23) * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3p5.py * Update src/transformers/models/gemma3p5/modular_gemma3p5.py Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Updating configs for the 2B variant in the conversion script * Using final generation config in conversion script --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Audio Integration (huggingface#12) * initial commit of Gemma 3n scaffold * Fixing param pass through on Gemm3nRMSNorm * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Adds AltUp to Gemma 3n * Adding Gemma 3n overall and text config with vision and audio config placeholders (huggingface#3) * Adding Gemma 3n text configs * Adding audio config placeholders * Adding a placeholder for vision configs * Updating MobileNetVisionConfig, inheriting TimmWrapperConfig * Updating text configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Removing altup configs to accept the suggested configs * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating altup config * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Addressing review comments and updating text configs * Adding a config for activation sparsity * Updating configs to pass through options to super class init and adjust some name prefixes * Updating laurel and altup with corrected config values * Normalizing sub_config initializers --------- Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Updating MLP with activation sparsity (huggingface#2) * Updating DecoderBlock for Gemma 3n (huggingface#3) * Initial Gemma3nTextModel (huggingface#4) NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference. * Adding KV Cache Sharing * Adds Einsum layer to Gemma 3n * Updating EinsumLayer API * Refactored kv cache sharing in attention * Adding KVStore for cache sharing * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update modular Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Update src/transformers/cache_utils.py Co-authored-by: Ryan Mullins <ryanmullins@google.com> * Undoing erroneous force push * Reverting RMSNorm to with_scale by default * Adds LAuReL to Gemma 3n * Updating KV Cache Sharing implementation * Updating the q and k norm definitions in the attention module * Fixing name error for q,k,v RMS norm to use the right 3n module * Updating MLP with activation sparsity * Updating DecoderBlock for Gemma 3n * Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code * Isolating KV Cache logic to relevant components * Fixing logic error in Gemma3nAttention.forward * Refactoring caching contributions and fixing kv_store initialization * Simplifying Configs * Remove errant self from super init call * Bug fix in the Attention module - changing self.head_dim to config.head_dim * Bug fixes in the LaurelBlock and RMS Norm super init call * removing redundant code from a merge * Adding per_layer_inputs to TextModel * Adding preprocess embeddings with altup * Adds per-layer-to-single output and a host of TODOs * Integrating altup predict with the model workflow and other minor bug fixes * Using nn.Embedding temporarily for text model * It goes forward * Minor refactor of attention sparsity and RoPE initialization * Fixing duplicate rope_scaling param bug when loading from pretrained --------- Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Normalizing on altup_num_inputs config option * Adding audio encoder config * Adds high-level components for Audio Encoder * Implement uniform reducer for Audio Encoder * Adding placeholders for Conformer components in Audio Encoder * Adding placeholders for SubSampleConvProjection components in Audio Encoder * Adding SequenceLayer component placeholders * Implementing Gemma3nAudioEncoder with nn.Sequential * Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential * Implementing Conformer model with SequenceLayers * Use OrderedDict in nn.Sequential initializers * Implements sl.Residual in Torch with nn.Sequential and OrderedDict * Adopting a base SequenceLayer class with default forward() method * Implementing sl.GatedLinearUnit in Torch * Implementing sl.Swish in Torch * Implementing sl.ReLU in Torch * Implementing sl.Scale in Torch * Removing sl.Dropout after tree-shaking * Implementing sl.RMSNorm in Torch with fake shape * Implementing sl.GroupNorm in Torch * Implementing sl.Conv2d in Torch * Implementing sl.Dense in Torch * Removing sl.Delay layers, which act as pass-throughs * Connecting shapes to configs in initializers * Removing sl.Emit * Implementing sl.ExpandDims in Torch * Adding sl.GradientClipping to Torch * Implementing sl.DenseShaped in Torch * Implementing sl.LDPA in Torch * Removing unused sl.CombinedQKVProj class * Fixing erroneous type hint * Implemnenting sl.DepthwiseConv1D in Torch * Implementing sl.MaskInvalid in Torch * Fixes for initialization * Fixes for saving weights * Removing einsums per feedback from HF staff * Removing Sequence Layers idioms from audio encoder * Fixes for reviewer comments * Converting sl.Frontend to FeatureExtractor * Updates for ConditionalGeneration.get_image_features * Adding a WIP draft of image_processing_gemma3n.py * Update modular Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> * Modular conversion after github suggested change * Text + image gives good results * Fixing image size preset * Draft of audio data in chat template * Removing image processing. Using SigLIP instead. * Audio input going end-to-end * Fixing dtype issues in audio encoder * x-lib formatting consistency * Adding example data * Save preprocessor_config.json from conversion script * Instrumentaiton for debugging * Additional instrumentation for preprocessing debugging * Updates to preprocessor, padding; produces correct end-to-end results on sample * Tackling configuraiton TODOs * Start of feature extractor refatcor * Adds Numpy version of USM extractor, removes Torch version and dependencies * Fixing AltUp.correct coef permute * Supporting batches of single audio segment inputs * Docstrings updates for config * In-lining audio feature extraction * Adjustments to conversion script and smoke test script --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: pculliton <phillipculliton@gmail.com> * Gemma 3n renaming * Removing test data and utilities * Renaming test files * Gemma 3n refactor * Fix tokenizer config in conversion script * Address reviewer feedback * FeatureExtractor returns float32 by default * Adding basic tests for audio, and input name for audio encoder * Audio integration test, updates to model_id for other integration tests * Use scales for q and k norms (huggingface#26) * Update audio integration test to use HF dataset * Reviewer feedback * Expand embedding table to full vocab size in weights conversion * Mix-n-match MatFormers for Gemma 3n (huggingface#25) * Remove in-place operations (huggingface#30) * chore: removing inplace ops * remove [tensor] * n pattern * chore: reviewer feedback in AudioEncoder and AltUp * More grad clipping * Dynamo compatibility * fix: cache slicing error * chore: simplify shared kv cache slicing * chore: vision encoder rename in timm * fix: image processor do_normalize=False * fixup: style * chore: model_doc * fix: docs for code quality * chore: repo consistency * fix: RMSNorm in float as in prior Gemmas * fix: per_layer_inputs = None * chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint * chore: repo consistency * Add initial unit tests for Gemma3nAudioFeatureExtractor (huggingface#27) * Add initial unit tests for Gemma3nAudioFeatureExtractor * Add basic unit tests for Gemma3nProcessor (huggingface#28) Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * parameterize tests --------- Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> * chore: code style * fix: test cases * style and consistency * fix config in the test to be coherent with layer cache sharing * fix hidden states in tests and code * inits and mappings * fix modality prefixes * test order and prefixes * fix test exception * fix class order and reduce model size for faster tests * restore _checkpoint_conversion_mapping to load Caual from Conditional * fix config mapping! * fix: reviewer feedback --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> * fix import test * add model args * auto_docstring * replace test path * consistency * skip tests for now * fix docstring for doc builder * skip unused attr --------- Co-authored-by: SindhuRaghuram97 <114270661+SindhuRaghuram97@users.noreply.github.com> Co-authored-by: Sindhu Raghuram <sindhuraghuram@google.com> Co-authored-by: raushan <raushan@huggingface.co> Co-authored-by: Mayank Chaturvedi <imayank@google.com> Co-authored-by: Douglas Reid <douglas-reid@users.noreply.github.com> Co-authored-by: Douglas Reid <21148125+douglas-reid@users.noreply.github.com> Co-authored-by: Xuan-Son Nguyen <thichthat@gmail.com> Co-authored-by: pculliton <phillipculliton@gmail.com> Co-authored-by: Aritra Roy Gosthipaty <aritra.born2fly@gmail.com> Co-authored-by: Cyril Vallez <cyril.vallez@gmail.com> Co-authored-by: Arthur <arthur.zucker@gmail.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
initial commit of Gemma 3n scaffold
Fixing param pass through on Gemm3p5RMSNorm
Adds Einsum layer to Gemma 3n
Updating EinsumLayer API
Undoing erroneous force push
Reverting RMSNorm to with_scale by default
Adds LAuReL to Gemma 3n
Adds AltUp to Gemma 3n
Adding Gemma3p5 overall and text config with vision and audio config placeholders (run_squad questions #3)
Adding gemma3p5 text configs
Adding audio config placeholders
Adding a placeholder for vision configs
Updating MobileNetVisionConfig, inheriting TimmWrapperConfig
Updating text configs
Update src/transformers/models/gemma3p5/modular_gemma3p5.py
Removing altup configs to accept the suggested configs
Update src/transformers/models/gemma3p5/modular_gemma3p5.py
Updating altup config
Update modular
Update modular
Update modular
Update modular
Addressing review comments and updating text configs
Adding a config for activation sparsity
Updating configs to pass through options to super class init and adjust some name prefixes
Updating laurel and altup with corrected config values
Normalizing sub_config initializers
Updating MLP with activation sparsity (Port tokenization for the multilingual model #2)
Updating DecoderBlock for Gemma 3n (run_squad questions #3)
Initial Gemm3nTextModel (Fix typo in subheader BertForQuestionAnswering #4)
NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference.
Adding KV Cache Sharing
Adds Einsum layer to Gemma 3n
Updating EinsumLayer API
Refactored kv cache sharing in attention
Adding KVStore for cache sharing
Update modular
Update modular
Update modular
Update src/transformers/cache_utils.py
Undoing erroneous force push
Reverting RMSNorm to with_scale by default
Adds LAuReL to Gemma 3n
Updating KV Cache Sharing implementation
Updating the q and k norm definitions in the attention module
Fixing name error for q,k,v RMS norm to use the right 3n module
Updating MLP with activation sparsity
Updating DecoderBlock for Gemma 3.5
Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code
Isolating KV Cache logic to relevant components
Fixing logic error in Gemma3nAttention.forward
Refactoring caching contributions and fixing kv_store initialization
Simplifying Configs
Remove errant self from super init call
Bug fix in the Attention module - changing self.head_dim to config.head_dim
Bug fixes in the LaurelBlock and RMS Norm super init call
removing redundant code from a merge
Adding per_layer_inputs to TextModel
Adding preprocess embeddings with altup
Adds per-layer-to-single output and a host of TODOs
Integrating altup predict with the model workflow and other minor bug fixes
Using nn.Embedding temporarily for text model
It goes forward
Minor refactor of attention sparsity and RoPE initialization
Fixing duplicate rope_scaling param bug when loading from pretrained
Normalizing on altup_num_inputs config option
regenerating modeling file after syncing to HEAD
Use torch.std(..., unbiased=False) for activation sparsity (fixed small typos in the README.md #8)
Refactoring to a single QVK Norm (Bug in run_classifier.py #13)
AltUp: support scale_corrected_output (fixed typo #14)
Converts einsums to nn.Linear (Develop #7)
Converts einsums to nn.Linear
Removing unused variables
Aligning SharedKVCache with HybridCache (Swapped to_seq_len/from_seq_len in comment #11)
Alinging SharedKVStore with HybridCache
Remove KVStore. Refactor apply_rotary_pos_emb for sharing
Addressing review comments
Supporting split modality embeddings in Gemma3n (Is there a plan to have a FP16 for GPU so to have larger batch size or longer text documents support ? #10)
Adding the Embedder class
Update modular
Update modular
Update modular
Update modular
Update modular
Update modular
Addressing review comments, adding audio embedding layers, integrating embedder with the remaining architecture, adding a forward method for conditional generation
Apply suggestions from code review
Update modular
Addressing review comments, prop drilling audio and vision configs to the text config
Removing TODO's that have been addressed
Simplify Embedder init and add audio embeddings
Embeddings refactor. Adds Gemma3nAudioEmbedder and Gemma3nVisionEmbedder
Refactoring vision and audio embeddings into ConditionalGeneration model
Updating attention mask for Gemma 3.5 (activation function in BERTIntermediate #15)
xxx_token_index to xxx_token_id
remvoing deprecated last_cache_position
Removing references to SigLIP
Always init per-layer inputs
Using torch.finfo().min for epsilon_tensor
Gemma3nDecoderLayer inherits from Gemma3DecoderLayer. Remove gating lambdas
fix modular GEMMA3N_INPUTS_DOCSTRING
Gemma3nAttention inherits from Gemma3Attention
Modular inheritance fixes
CausalLM conversion script for 4B model (Excluding AdamWeightDecayOptimizer internal variables from restoring #16)
Add Gemma3n Audio Encoder (Failure during pytest (and solution for python3) #6)
initial commit of Gemma 3.5 scaffold
Fixing param pass through on Gemm3nRMSNorm
Adds Einsum layer to Gemma 3.5
Updating EinsumLayer API
Undoing erroneous force push
Reverting RMSNorm to with_scale by default
Adds LAuReL to Gemma 3n
Adds AltUp to Gemma 3n
Adding Gemma3n overall and text config with vision and audio config placeholders (run_squad questions #3)
Adding gemma3n text configs
Adding audio config placeholders
Adding a placeholder for vision configs
Updating MobileNetVisionConfig, inheriting TimmWrapperConfig
Updating text configs
Update modular
Removing altup configs to accept the suggested configs
Update modular
Updating altup config
Update modular
Update modular
Update modular
Update modular
Addressing review comments and updating text configs
Adding a config for activation sparsity
Updating configs to pass through options to super class init and adjust some name prefixes
Updating laurel and altup with corrected config values
Normalizing sub_config initializers
Updating MLP with activation sparsity (Port tokenization for the multilingual model #2)
Updating DecoderBlock for Gemma 3.5 (run_squad questions #3)
Initial Gemm3nTextModel (Fix typo in subheader BertForQuestionAnswering #4)
NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference.
Adding KV Cache Sharing
Adds Einsum layer to Gemma 3.5
Updating EinsumLayer API
Refactored kv cache sharing in attention
Adding KVStore for cache sharing
Update modular
Update modular
Update modular
Update src/transformers/cache_utils.py
Undoing erroneous force push
Reverting RMSNorm to with_scale by default
Adds LAuReL to Gemma 3n
Updating KV Cache Sharing implementation
Updating the q and k norm definitions in the attention module
Fixing name error for q,k,v RMS norm to use the right Gemma 3n module
Updating MLP with activation sparsity
Updating DecoderBlock for Gemma 3.5
Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code
Isolating KV Cache logic to relevant components
Fixing logic error in Gemma3nAttention.forward
Refactoring caching contributions and fixing kv_store initialization
Simplifying Configs
Remove errant self from super init call
Bug fix in the Attention module - changing self.head_dim to config.head_dim
Bug fixes in the LaurelBlock and RMS Norm super init call
removing redundant code from a merge
Adding per_layer_inputs to TextModel
Adding preprocess embeddings with altup
Adds per-layer-to-single output and a host of TODOs
Integrating altup predict with the model workflow and other minor bug fixes
Using nn.Embedding temporarily for text model
It goes forward
Minor refactor of attention sparsity and RoPE initialization
Fixing duplicate rope_scaling param bug when loading from pretrained
Normalizing on altup_num_inputs config option
Adding audio encoder config
Adds high-level components for Audio Encoder
Implement uniform reducer for Audio Encoder
Adding placeholders for Conformer components in Audio Encoder
Adding placeholders for SubSampleConvProjection components in Audio Encoder
Adding SequenceLayer component placeholders
Implementing Gemma3nAudioEncoder with nn.Sequential
Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential
Implementing Conformer model with SequenceLayers
Use OrderedDict in nn.Sequential initializers
Implements sl.Residual in Torch with nn.Sequential and OrderedDict
Adopting a base SequenceLayer class with default forward() method
Implementing sl.GatedLinearUnit in Torch
Implementing sl.Swish in Torch
Implementing sl.ReLU in Torch
Implementing sl.Scale in Torch
Removing sl.Dropout after tree-shaking
Implementing sl.RMSNorm in Torch with fake shape
Implementing sl.GroupNorm in Torch
Implementing sl.Conv2d in Torch
Implementing sl.Dense in Torch
Removing sl.Delay layers, which act as pass-throughs
Connecting shapes to configs in initializers
Removing sl.Emit
Implementing sl.ExpandDims in Torch
Adding sl.GradientClipping to Torch
Implementing sl.DenseShaped in Torch
Implementing sl.LDPA in Torch
Removing unused sl.CombinedQKVProj class
Fixing erroneous type hint
Implemnenting sl.DepthwiseConv1D in Torch
Implementing sl.MaskInvalid in Torch
Fixes for initialization
Fixes for saving weights
Removing einsums per feedback from HF staff
Removing Sequence Layers idioms from audio encoder
Fixes for reviewer comments
CausalLM conversion script for 4B model
inv_timescales to non-persistent buffer
Addressing audio encoder Attention feedback
Addressing Gemma3nAudioSSCPConvBlock feedback
Addressing Gemma3nAudioConformerAttention feedback
Addressing padding feedback
Weights conversion loads audio state dict
Always use vision_config so saving works
Token id updates for configs
Stubs for interleaving audio embs
Addressing reviewer feedback
Fixing cache access error
Removing duplicate code from a bad merge
Gemma 3n Text + Vision Part 1 (activation function in BERTIntermediate #17)
testing utilities for numerics comparisons
Corrected einsum to nn.Linear weights conversion
Inherit scaled word embs from Gemma3 not Bart
Fixing transposes for collapsed linears
More transpose fixes
numpy api fix
RMSNorm: Explicit kwargs, scale_shift=0.0 when with_scale=True
Force AltUp to float32
Updating debugging script for AudioEncoder debugging
Support divide_weight_by_sqrt_fan_in from JAX for per-layer inputs
Correcting attention einsum conversions
RMSNorm in type of x
Fixing douplicate laurel norm/gating
KV sharing using the right previous indices
Refactor kv shared index computation. Correct frac_shared_layers
Use num_shared_layers instead of inferring from a fraction
fixing a bug for logging
Fix shared data_ptrs in altup inits
rope: adjust proj -> norm -> rope to preserve computation (model loading the checkpoint error #20)
rope: adjust proj -> norm -> rope to preserve computation
Removing some breaking language model fluff in ConditionalGeneration
Consolidate query_states transforms
Vectorize the loops in AltUp (will you push the pytorch code for the pre-training process? #19)
Vectorize the loops in AltUp
fix typo
Expanding to support batched inputs
remove extra debug script
Fix AltUp.forward
Add 'scale_shift=0.0, with_scale=True' to the final norm in TextModel
Convert norm to 1/sqrt (Fix some glitches in extract_features.py #21)
Convert norm to 1/sqrt
Scale shift change per Phil's rec
Adding default activation sparsity
Fixing 2B config in weights conversion script
Fixing RMSNorm parameters - adding scale_shift and with_scale
Correcting query pre-attention scaling
Adding query_rescale_scalar to text config
Adding layer_idx to MLP
Permafix for input_layernorm
Use 1/sqrt instead of rsqrt in DecoderLayer
Fix o_proj conversion
Conversion script update for vision encoder
Removing logging for debugging timm model
Fixing bugs in Gemma3nForConditionalGeneration for text generation
Generating the modeling_gemma3n.py file
Removing the addition of an erroneous line in the modeling file
Adding gemma3n text model to modeling_auto
Bugfix: Updating the interleaving of inputs_embeds and vision_embeds
Updating the modeling file with the latest bugfix changes
Updating models/auto for Gemma 3n
using AutoTokenizer in forward test
Adding processing_gemma3n.py
Gemma 3n configured for AutoModel. Conversion script updated.
Removing errant merge artifacts
Removing errant debugging statements from Gemma 3
Gemma3n audio model (include the output layer in the model using the pretrained weights #18)
testing utilities for numerics comparisons
Implement CumulativeGroupNorm and add to SubSampleConvProjection and SSCPConvBlock
Add audio version of forward script based on RyanMullins' implementation
Updating to match encoder tests. WIP: config question needs resolving
Updates to audio classes to enable end-to-end running
Removing vestigial classes, cleaning up print statements
Adding SiLU / Swish to audio conformer feed forward block
Shifted Gemma3p5Audio naming prefix to Gemma3NanoAudio
Adding outputs to audio test
Fixes to padding in SSCP and 1D convolution, align RMS Norm with wider model
Update forward test to load from local weights
Update conversion to process / output audio layers
Update all to export audio encoder
AutoModel registration for Gemma 3n Audio
Use AutoModel for ConditionalGeneration.audio_tower
Fixing input_proj_linear transpose
Fixing Gemma3NanoAudioConformerAttention.post conversion
Fixing Gemma3NanoAudioSSCPConvBlock.conv weights conversion
Correcting indentation issue on Gemma3p5RMSNorm
Text + Vision Part 2 (ValueError while using --optimize_on_cpu #23)
Updates for ConditionalGeneration.get_image_features
Adding a WIP draft of image_processing_gemma3p5.py
Update src/transformers/models/gemma3p5/modular_gemma3p5.py
Modular conversion after github suggested change
Text + image gives good results
Fixing image size preset
Updating configs for the 2B variant in the conversion script
Using final generation config in conversion script
Audio Integration (py2 code #12)
initial commit of Gemma 3n scaffold
Fixing param pass through on Gemm3nRMSNorm
Adds Einsum layer to Gemma 3n
Updating EinsumLayer API
Undoing erroneous force push
Reverting RMSNorm to with_scale by default
Adds LAuReL to Gemma 3n
Adds AltUp to Gemma 3n
Adding Gemma 3n overall and text config with vision and audio config placeholders (run_squad questions #3)
Adding Gemma 3n text configs
Adding audio config placeholders
Adding a placeholder for vision configs
Updating MobileNetVisionConfig, inheriting TimmWrapperConfig
Updating text configs
Update modular
Removing altup configs to accept the suggested configs
Update modular
Updating altup config
Update modular
Update modular
Update modular
Update modular
Addressing review comments and updating text configs
Adding a config for activation sparsity
Updating configs to pass through options to super class init and adjust some name prefixes
Updating laurel and altup with corrected config values
Normalizing sub_config initializers
Updating MLP with activation sparsity (Port tokenization for the multilingual model #2)
Updating DecoderBlock for Gemma 3n (run_squad questions #3)
Initial Gemma3nTextModel (Fix typo in subheader BertForQuestionAnswering #4)
NOTE: This implementation WILL CHANGE in the coming weeks, however, changes will be strictly additive and this will remain a suitable baseline for downstream implementations to reference.
Adding KV Cache Sharing
Adds Einsum layer to Gemma 3n
Updating EinsumLayer API
Refactored kv cache sharing in attention
Adding KVStore for cache sharing
Update modular
Update modular
Update modular
Update src/transformers/cache_utils.py
Undoing erroneous force push
Reverting RMSNorm to with_scale by default
Adds LAuReL to Gemma 3n
Updating KV Cache Sharing implementation
Updating the q and k norm definitions in the attention module
Fixing name error for q,k,v RMS norm to use the right 3n module
Updating MLP with activation sparsity
Updating DecoderBlock for Gemma 3n
Updating kv cache sharing implementation with the use of a cache buffer and refactoring some lines of code
Isolating KV Cache logic to relevant components
Fixing logic error in Gemma3nAttention.forward
Refactoring caching contributions and fixing kv_store initialization
Simplifying Configs
Remove errant self from super init call
Bug fix in the Attention module - changing self.head_dim to config.head_dim
Bug fixes in the LaurelBlock and RMS Norm super init call
removing redundant code from a merge
Adding per_layer_inputs to TextModel
Adding preprocess embeddings with altup
Adds per-layer-to-single output and a host of TODOs
Integrating altup predict with the model workflow and other minor bug fixes
Using nn.Embedding temporarily for text model
It goes forward
Minor refactor of attention sparsity and RoPE initialization
Fixing duplicate rope_scaling param bug when loading from pretrained
Normalizing on altup_num_inputs config option
Adding audio encoder config
Adds high-level components for Audio Encoder
Implement uniform reducer for Audio Encoder
Adding placeholders for Conformer components in Audio Encoder
Adding placeholders for SubSampleConvProjection components in Audio Encoder
Adding SequenceLayer component placeholders
Implementing Gemma3nAudioEncoder with nn.Sequential
Implementing Gemma3nAudioSubSampleConvProjection with nn.Sequential
Implementing Conformer model with SequenceLayers
Use OrderedDict in nn.Sequential initializers
Implements sl.Residual in Torch with nn.Sequential and OrderedDict
Adopting a base SequenceLayer class with default forward() method
Implementing sl.GatedLinearUnit in Torch
Implementing sl.Swish in Torch
Implementing sl.ReLU in Torch
Implementing sl.Scale in Torch
Removing sl.Dropout after tree-shaking
Implementing sl.RMSNorm in Torch with fake shape
Implementing sl.GroupNorm in Torch
Implementing sl.Conv2d in Torch
Implementing sl.Dense in Torch
Removing sl.Delay layers, which act as pass-throughs
Connecting shapes to configs in initializers
Removing sl.Emit
Implementing sl.ExpandDims in Torch
Adding sl.GradientClipping to Torch
Implementing sl.DenseShaped in Torch
Implementing sl.LDPA in Torch
Removing unused sl.CombinedQKVProj class
Fixing erroneous type hint
Implemnenting sl.DepthwiseConv1D in Torch
Implementing sl.MaskInvalid in Torch
Fixes for initialization
Fixes for saving weights
Removing einsums per feedback from HF staff
Removing Sequence Layers idioms from audio encoder
Fixes for reviewer comments
Converting sl.Frontend to FeatureExtractor
Updates for ConditionalGeneration.get_image_features
Adding a WIP draft of image_processing_gemma3n.py
Update modular
Modular conversion after github suggested change
Text + image gives good results
Fixing image size preset
Draft of audio data in chat template
Removing image processing. Using SigLIP instead.
Audio input going end-to-end
Fixing dtype issues in audio encoder
x-lib formatting consistency
Adding example data
Save preprocessor_config.json from conversion script
Instrumentaiton for debugging
Additional instrumentation for preprocessing debugging
Updates to preprocessor, padding; produces correct end-to-end results on sample
Tackling configuraiton TODOs
Start of feature extractor refatcor
Adds Numpy version of USM extractor, removes Torch version and dependencies
Fixing AltUp.correct coef permute
Supporting batches of single audio segment inputs
Docstrings updates for config
In-lining audio feature extraction
Adjustments to conversion script and smoke test script
Gemma 3n renaming
Removing test data and utilities
Renaming test files
Gemma 3n refactor
Fix tokenizer config in conversion script
Address reviewer feedback
FeatureExtractor returns float32 by default
Adding basic tests for audio, and input name for audio encoder
Audio integration test, updates to model_id for other integration tests
Use scales for q and k norms (Checkpoints not saved #26)
Update audio integration test to use HF dataset
Reviewer feedback
Expand embedding table to full vocab size in weights conversion
Mix-n-match MatFormers for Gemma 3n (can you push the run-pretraining and create_pretraining_data codes? #25)
Remove in-place operations ([Feature request] Add example of finetuning the pretrained models on custom corpus #30)
chore: removing inplace ops
remove [tensor] * n pattern
chore: reviewer feedback in AudioEncoder and AltUp
More grad clipping
Dynamo compatibility
fix: cache slicing error
chore: simplify shared kv cache slicing
chore: vision encoder rename in timm
fix: image processor do_normalize=False
fixup: style
chore: model_doc
fix: docs for code quality
chore: repo consistency
fix: RMSNorm in float as in prior Gemmas
fix: per_layer_inputs = None
chore: Gemma3nForCausalLM from Gemma3nForConditionalGeneration checkpoint
chore: repo consistency
Add initial unit tests for Gemma3nAudioFeatureExtractor (how to load checkpoint? #27)
Add initial unit tests for Gemma3nAudioFeatureExtractor
Add basic unit tests for Gemma3nProcessor (speed is very slow #28)
parameterize tests
chore: code style
fix: test cases
style and consistency
fix config in the test to be coherent with layer cache sharing
fix hidden states in tests and code
inits and mappings
fix modality prefixes
test order and prefixes
fix test exception
fix class order and reduce model size for faster tests
restore _checkpoint_conversion_mapping to load Caual from Conditional
fix config mapping!
fix: reviewer feedback
What does this PR do?
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.