load draft model fix #7506
load draft model fix #7506
Conversation
There was a problem hiding this comment.
Summary of Changes
Hello @yilian49, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request implements a crucial fix for loading specific model architectures, particularly the DeepseekV3ForCausalLMNextN model. By introducing a conditional call to the model's post-loading weight processing, it ensures that these models are initialized with the correct parameters, resolving potential loading issues and enabling their proper functionality within the system.
Highlights
- Model Loading Logic: Introduced conditional logic within the
load_modelfunction to specifically handle theDeepseekV3ForCausalLMNextNmodel architecture. When this architecture is detected, themodel.post_load_weights()method is now called with the argumentis_nextn=True. - Targeted Fix: This change is a targeted fix, likely addressing an issue with the post-processing of weights for the
DeepseekV3ForCausalLMNextNmodel, ensuring it loads correctly as a 'draft model'.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces a fix for loading a draft model, specifically for DeepseekV3ForCausalLMNextN within the DummyModelLoader. The change adds special handling during the post-weight-loading process for this architecture. I've suggested refactoring to improve maintainability.
| if model_config.hf_config.architectures[0] == "DeepseekV3ForCausalLMNextN": | ||
| model.post_load_weights(is_nextn = True) | ||
| else: | ||
| model.post_load_weights() |
There was a problem hiding this comment.
Consider simplifying this conditional logic. Instead of checking the architecture string directly, pass is_nextn as an argument to post_load_weights based on a boolean flag1. This moves the architecture-specific logic into the DeepseekV3ForCausalLMNextN class itself, improving maintainability.
is_nextn = model_config.hf_config.architectures[0] == "DeepseekV3ForCausalLMNextN"
model.post_load_weights(is_nextn=is_nextn)Style Guide References
Footnotes
|
@yilian49 Could you please fix the linter issue? |
|
@ch-wan Just fixed |
Motivation
When loading deepseek v3, post_load_weights function called in loader.py does not pass in a is_nextn argument which reports the following error when loading DeepseekModelNextN models.
Modifications
Add a check for if model architecture is "DeepseekV3ForCausalLMNextN" and pass in is_nextn = true in model.post_load_weights.
Note
There could be a more elegant solution.
Update: This error only happens when using flag "--load-format dummy".
Command
SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=256 python3 -m sglang.launch_server
--attention-backend fa3
--model-path /mnt/shared-fs/models/deepseek-ai/DeepSeek-V3-0324
--tp 8
--nnodes 1
--node-rank 0
--trust-remote-code
--max-running-requests 32
--mem-fraction-static 0.7 \
--chunked-prefill-size 2048
--schedule-policy fcfs \
--speculative-algo NEXTN
--speculative-num-steps 4
--speculative-eagle-topk 1
--speculative-num-draft-tokens 4
--cuda-graph-max-bs 32 \
--disable-shared-experts-fusion
--load-format dummy
Error pattern
[2025-06-24 23:28:08 TP1] Scheduler hit an exception: Traceback (most recent call last):
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/managers/scheduler.py", line 2631, in run_scheduler_process
scheduler = Scheduler(server_args, port_args, gpu_id, tp_rank, pp_rank, dp_rank)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/managers/scheduler.py", line 328, in init
self.draft_worker = EAGLEWorker(
^^^^^^^^^^^^
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/speculative/eagle_worker.py", line 124, in init
super().init(
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/managers/tp_worker.py", line 79, in init
self.model_runner = ModelRunner(
^^^^^^^^^^^^
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/model_executor/model_runner.py", line 220, in init
self.initialize(min_per_gpu_memory)
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/model_executor/model_runner.py", line 262, in initialize
self.load_model()
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/model_executor/model_runner.py", line 561, in load_model
self.model = get_model(
^^^^^^^^^^
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/model_loader/init.py", line 22, in get_model
return loader.load_model(
^^^^^^^^^^^^^^^^^^
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/model_loader/loader.py", line 558, in load_model
model.post_load_weights()
File "/mnt/shared-fs/yilian/sglang/python/sglang/srt/models/deepseek_v2.py", line 1844, in post_load_weights
self.model.layers[layer_id].self_attn
^^^^^^^^^^^^^^^^^
File "/mnt/shared-fs/yilian/sgl-yilian/lib/python3.11/site-packages/torch/nn/modules/module.py", line 1940, in getattr
raise AttributeError(
AttributeError: 'DeepseekModelNextN' object has no attribute 'layers'