feat: prefill aware routing #1895

PeaBrane · 2025-07-11T20:54:08Z

Overview:

To better predict routing latency, we need to consider both new and in-progress prefill tokens—not just overlaps, which only reflect new tokens.

Core changes are in sequence.rs and scheduler.rs. Brief algorithmic overview is:

When a request comes in, mark it as prefill immediately, and increment the new tokens to active_tokens count
Upon the first token generation, mark the request as decode, and reduce the token count for that request to 1 (assuming pure auto-regression)
On request completion, reduce token count to 0

Results:

Testing setups:

vllm v1 engines: 8 workers, 16384 batched tokens, block size 64, L40 GPUs
router: temperature 0, overlap weight 1, listening to KV events
requests: 200 requests with 20 prefix prompts, with prefix length increasing to control data hit rate. ISL 14k and OSL 200. Concurrency 20.

500 GPU blocks (intentional starvation) to simulate low VRAM settings

No block starvation (vllm default dynamic num blocks)

Summary by CodeRabbit

New Features
- Improved scheduling logic now accounts for both token-level and block-level metrics, enhancing resource management and worker selection.
- More precise tracking of active and potential tokens, leading to better scheduling accuracy.
Refactor
- Updated various internal data structures to support token-level granularity.
- Renamed and added fields to improve clarity and consistency in scheduling and tracking.
Bug Fixes
- Ensured prompt handling of initial token pushes in streaming responses.

coderabbitai · 2025-07-11T21:24:24Z

Walkthrough

This change refactors the KV router and scheduler to track both token-level and block-level resource usage for more granular scheduling decisions. It updates data structures, scheduling logic, and related method signatures to account for active and potential tokens in addition to blocks, and standardizes the type for the overlap_blocks field.

Changes

File(s)	Change Summary
`components/metrics/src/bin/mock_worker.rs`, `components/metrics/src/main.rs`	Explicitly cast `overlap_blocks` to appropriate types (`u32` and `usize`) during struct construction and method calls.
`lib/llm/src/kv_router/protocols.rs`	Changed `overlap_blocks` field in `WorkerSelectionResult` from `usize` to `u32`.
`lib/llm/src/kv_router.rs`	Added `first_push_done` flag to control initial token push in streaming responses.
`lib/llm/src/kv_router/scheduler.rs`	Incorporated token-level metrics into scheduling; updated data structures, logic, and formulas; added/renamed fields and methods.
`lib/llm/src/kv_router/sequence.rs`	Added tracking for active/prefill tokens; updated/add methods for token accounting; adjusted sequence management and tests.

Sequence Diagram(s)

sequenceDiagram
    participant Client
    participant Scheduler
    participant Sequences
    participant WorkerSelector

    Client->>Scheduler: schedule(request_id, tokens, block_size, overlaps)
    Scheduler->>Sequences: potential_blocks_and_tokens(token_sequence, overlaps)
    Sequences-->>Scheduler: (potential_blocks, potential_tokens)
    Scheduler->>WorkerSelector: select_worker(workers, request, block_size)
    WorkerSelector-->>Scheduler: WorkerSelectionResult (overlap_blocks: u32)
    Scheduler-->>Client: SchedulingResponse (overlap_blocks, worker_id)

Possibly related PRs

ai-dynamo/dynamo#1848: Modifies token push handling in the KV router and scheduler, directly relating to changes in token/block push logic.
ai-dynamo/dynamo#1731: Refactors and introduces the overlap_blocks field and associated scheduling logic, closely connected to the main PR’s updates.

Poem

In the warren where tokens hop and blocks align,
Rabbits count each nibble, not just by design.
With overlap blocks now tidy and tokens in tow,
The scheduler’s smarter—oh, how numbers grow!
From fields to flows, we keep it precise,
For every byte and bunny, a system that’s nice.
🐇✨

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

🪧 Tips

Chat

There are 3 ways to chat with CodeRabbit:

Review comments: Directly reply to a review comment made by CodeRabbit. Example:
- I pushed a fix in commit <commit_id>, please review it.
- Explain this complex logic.
- Open a follow-up GitHub issue for this discussion.
Files and specific lines of code (under the "Files changed" tab): Tag @coderabbitai in a new review comment at the desired location with your query. Examples:
- @coderabbitai explain this code block.
- @coderabbitai modularize this function.
PR comments: Tag @coderabbitai in a new PR comment to ask questions about the PR branch. For the best results, please provide a very specific query, as very limited context is provided in this mode. Examples:
- @coderabbitai gather interesting stats about this repository and render them as a table. Additionally, render a pie chart showing the language distribution in the codebase.
- @coderabbitai read src/utils.ts and explain its main purpose.
- @coderabbitai read the files in the src/scheduler package and generate a class diagram using mermaid and a README in the markdown format.
- @coderabbitai help me debug CodeRabbit configuration file.

Support

Need help? Create a ticket on our support page for assistance with any issues or questions.

Note: Be mindful of the bot's finite context window. It's strongly recommended to break down tasks such as reading entire modules into smaller chunks. For a focused discussion, use review comments to chat about specific files and their changes, instead of using the PR comments.

CodeRabbit Commands (Invoked using PR comments)

@coderabbitai pause to pause the reviews on a PR.
@coderabbitai resume to resume the paused reviews.
@coderabbitai review to trigger an incremental review. This is useful when automatic reviews are disabled for the repository.
@coderabbitai full review to do a full review from scratch and review all the files again.
@coderabbitai summary to regenerate the summary of the PR.
@coderabbitai generate sequence diagram to generate a sequence diagram of the changes in this PR.
@coderabbitai resolve resolve all the CodeRabbit review comments.
@coderabbitai configuration to show the current CodeRabbit configuration for the repository.
@coderabbitai help to get help.

Other keywords and placeholders

Add @coderabbitai ignore anywhere in the PR description to prevent this PR from being reviewed.
Add @coderabbitai summary to generate the high-level summary at a specific location in the PR description.
Add @coderabbitai anywhere in the PR title to generate the title automatically.

Documentation and Community

Visit our Documentation for detailed information on how to use CodeRabbit.
Join our Discord Community to get help, request features, and share feedback.
Follow us on X/Twitter for updates and announcements.

coderabbitai

Actionable comments posted: 2

📜 Review details

Configuration used: .coderabbit.yaml
Review profile: CHILL
Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 35c5606 and 366c175.

📒 Files selected for processing (6)

components/metrics/src/bin/mock_worker.rs (1 hunks)
components/metrics/src/main.rs (1 hunks)
lib/llm/src/kv_router.rs (2 hunks)
lib/llm/src/kv_router/protocols.rs (1 hunks)
lib/llm/src/kv_router/scheduler.rs (7 hunks)
lib/llm/src/kv_router/sequence.rs (17 hunks)

🧰 Additional context used

🧠 Learnings (7)

📓 Common learnings

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1285
File: lib/llm/src/kv_router/scheduler.rs:260-266
Timestamp: 2025-05-30T06:34:12.785Z
Learning: In the KV router scheduler code, PeaBrane prefers fail-fast behavior over silent failure handling. When accessing worker metrics data that could be out-of-bounds (like dp_rank indexing), explicit panics are preferred over graceful degradation with continue statements to ensure data integrity issues are caught early.

components/metrics/src/main.rs (1)

Learnt from: ryanolson
PR: ai-dynamo/dynamo#1093
File: lib/llm/src/block_manager/block/registry.rs:98-122
Timestamp: 2025-05-29T06:20:12.901Z
Learning: In lib/llm/src/block_manager/block/registry.rs, the background task spawned for handling unregister notifications uses detached concurrency by design. The JoinHandle is intentionally not stored as this represents a reasonable architectural tradeoff for a long-running cleanup task.

components/metrics/src/bin/mock_worker.rs (4)

Learnt from: alec-flowers
PR: ai-dynamo/dynamo#1181
File: lib/llm/src/kv_router/publisher.rs:379-425
Timestamp: 2025-05-29T00:02:35.018Z
Learning: In lib/llm/src/kv_router/publisher.rs, the functions `create_stored_blocks` and `create_stored_block_from_parts` are correctly implemented and not problematic duplications of existing functionality elsewhere in the codebase.

Learnt from: jthomson04
PR: ai-dynamo/dynamo#1429
File: lib/runtime/src/utils/leader_worker_barrier.rs:69-72
Timestamp: 2025-06-08T03:12:03.985Z
Learning: In the leader-worker barrier implementation in lib/runtime/src/utils/leader_worker_barrier.rs, the `wait_for_key_count` function correctly uses exact equality (`==`) instead of greater-than-or-equal (`>=`) because worker IDs must be unique (enforced by etcd create-only operations), ensuring exactly the expected number of workers can register.

Learnt from: kthui
PR: ai-dynamo/dynamo#1424
File: lib/runtime/src/pipeline/network/egress/push_router.rs:204-209
Timestamp: 2025-06-13T22:07:24.843Z
Learning: The codebase uses async-nats version 0.40, not the older nats crate. Error handling should use async_nats::error::Error variants, not nats::Error variants.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1392
File: lib/llm/src/kv_router/scoring.rs:35-46
Timestamp: 2025-06-05T01:02:15.318Z
Learning: In lib/llm/src/kv_router/scoring.rs, PeaBrane prefers panic-based early failure over Result-based error handling for the worker_id() method to catch invalid data early during development.

lib/llm/src/kv_router/protocols.rs (5)

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1392
File: lib/llm/src/kv_router/scoring.rs:35-46
Timestamp: 2025-06-05T01:02:15.318Z
Learning: In lib/llm/src/kv_router/scoring.rs, PeaBrane prefers panic-based early failure over Result-based error handling for the worker_id() method to catch invalid data early during development.

Learnt from: alec-flowers
PR: ai-dynamo/dynamo#1181
File: lib/llm/src/kv_router/publisher.rs:379-425
Timestamp: 2025-05-29T00:02:35.018Z
Learning: In lib/llm/src/kv_router/publisher.rs, the functions `create_stored_blocks` and `create_stored_block_from_parts` are correctly implemented and not problematic duplications of existing functionality elsewhere in the codebase.

Learnt from: jthomson04
PR: ai-dynamo/dynamo#1429
File: lib/runtime/src/utils/leader_worker_barrier.rs:69-72
Timestamp: 2025-06-08T03:12:03.985Z
Learning: In the leader-worker barrier implementation in lib/runtime/src/utils/leader_worker_barrier.rs, the `wait_for_key_count` function correctly uses exact equality (`==`) instead of greater-than-or-equal (`>=`) because worker IDs must be unique (enforced by etcd create-only operations), ensuring exactly the expected number of workers can register.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1285
File: lib/llm/src/kv_router/scoring.rs:58-63
Timestamp: 2025-05-30T06:38:09.630Z
Learning: In lib/llm/src/kv_router/scoring.rs, the user prefers to keep the panic behavior when calculating load_avg and variance with empty endpoints rather than adding guards for division by zero. They want the code to fail fast on this error condition.

Learnt from: ryanolson
PR: ai-dynamo/dynamo#1093
File: lib/llm/src/block_manager/block/registry.rs:98-122
Timestamp: 2025-05-29T06:20:12.901Z
Learning: In lib/llm/src/block_manager/block/registry.rs, the background task spawned for handling unregister notifications uses detached concurrency by design. The JoinHandle is intentionally not stored as this represents a reasonable architectural tradeoff for a long-running cleanup task.

lib/llm/src/kv_router.rs (5)

Learnt from: alec-flowers
PR: ai-dynamo/dynamo#1181
File: lib/llm/src/kv_router/publisher.rs:379-425
Timestamp: 2025-05-29T00:02:35.018Z
Learning: In lib/llm/src/kv_router/publisher.rs, the functions `create_stored_blocks` and `create_stored_block_from_parts` are correctly implemented and not problematic duplications of existing functionality elsewhere in the codebase.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1285
File: lib/llm/src/kv_router/scoring.rs:58-63
Timestamp: 2025-05-30T06:38:09.630Z
Learning: In lib/llm/src/kv_router/scoring.rs, the user prefers to keep the panic behavior when calculating load_avg and variance with empty endpoints rather than adding guards for division by zero. They want the code to fail fast on this error condition.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1392
File: lib/llm/src/kv_router/scoring.rs:35-46
Timestamp: 2025-06-05T01:02:15.318Z
Learning: In lib/llm/src/kv_router/scoring.rs, PeaBrane prefers panic-based early failure over Result-based error handling for the worker_id() method to catch invalid data early during development.

Learnt from: ishandhanani
PR: ai-dynamo/dynamo#1626
File: lib/llm/src/preprocessor.rs:238-239
Timestamp: 2025-06-24T20:59:35.725Z
Learning: In lib/llm/src/preprocessor.rs, the `sampling_options` call in the `preprocess_request` method is placed in the common section after the match statement on `request.prompt_input_type()`, meaning it applies to both `PromptInput::Tokens` and `PromptInput::Text` request types.

Learnt from: t-ob
PR: ai-dynamo/dynamo#1290
File: launch/dynamo-run/src/subprocess/sglang_inc.py:80-110
Timestamp: 2025-06-03T10:17:51.711Z
Learning: The sglang `async_encode` method does not support streaming options, so collecting all embeddings before yielding is the correct approach for embedding requests.

lib/llm/src/kv_router/scheduler.rs (6)

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1285
File: lib/llm/src/kv_router/scheduler.rs:260-266
Timestamp: 2025-05-30T06:34:12.785Z
Learning: In the KV router scheduler code, PeaBrane prefers fail-fast behavior over silent failure handling. When accessing worker metrics data that could be out-of-bounds (like dp_rank indexing), explicit panics are preferred over graceful degradation with continue statements to ensure data integrity issues are caught early.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1285
File: lib/llm/src/kv_router/scoring.rs:58-63
Timestamp: 2025-05-30T06:38:09.630Z
Learning: In lib/llm/src/kv_router/scoring.rs, the user prefers to keep the panic behavior when calculating load_avg and variance with empty endpoints rather than adding guards for division by zero. They want the code to fail fast on this error condition.

Learnt from: alec-flowers
PR: ai-dynamo/dynamo#1181
File: lib/llm/src/kv_router/publisher.rs:379-425
Timestamp: 2025-05-29T00:02:35.018Z
Learning: In lib/llm/src/kv_router/publisher.rs, the functions `create_stored_blocks` and `create_stored_block_from_parts` are correctly implemented and not problematic duplications of existing functionality elsewhere in the codebase.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1392
File: lib/llm/src/kv_router/scoring.rs:35-46
Timestamp: 2025-06-05T01:02:15.318Z
Learning: In lib/llm/src/kv_router/scoring.rs, PeaBrane prefers panic-based early failure over Result-based error handling for the worker_id() method to catch invalid data early during development.

Learnt from: jthomson04
PR: ai-dynamo/dynamo#1429
File: lib/runtime/src/utils/leader_worker_barrier.rs:69-72
Timestamp: 2025-06-08T03:12:03.985Z
Learning: In the leader-worker barrier implementation in lib/runtime/src/utils/leader_worker_barrier.rs, the `wait_for_key_count` function correctly uses exact equality (`==`) instead of greater-than-or-equal (`>=`) because worker IDs must be unique (enforced by etcd create-only operations), ensuring exactly the expected number of workers can register.

Learnt from: ryanolson
PR: ai-dynamo/dynamo#1093
File: lib/llm/src/block_manager/block/registry.rs:98-122
Timestamp: 2025-05-29T06:20:12.901Z
Learning: In lib/llm/src/block_manager/block/registry.rs, the background task spawned for handling unregister notifications uses detached concurrency by design. The JoinHandle is intentionally not stored as this represents a reasonable architectural tradeoff for a long-running cleanup task.

lib/llm/src/kv_router/sequence.rs (5)

Learnt from: alec-flowers
PR: ai-dynamo/dynamo#1181
File: lib/llm/src/kv_router/publisher.rs:379-425
Timestamp: 2025-05-29T00:02:35.018Z
Learning: In lib/llm/src/kv_router/publisher.rs, the functions `create_stored_blocks` and `create_stored_block_from_parts` are correctly implemented and not problematic duplications of existing functionality elsewhere in the codebase.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1392
File: lib/llm/src/kv_router/scoring.rs:35-46
Timestamp: 2025-06-05T01:02:15.318Z
Learning: In lib/llm/src/kv_router/scoring.rs, PeaBrane prefers panic-based early failure over Result-based error handling for the worker_id() method to catch invalid data early during development.

Learnt from: ryanolson
PR: ai-dynamo/dynamo#1093
File: lib/llm/src/block_manager/block/registry.rs:98-122
Timestamp: 2025-05-29T06:20:12.901Z
Learning: In lib/llm/src/block_manager/block/registry.rs, the background task spawned for handling unregister notifications uses detached concurrency by design. The JoinHandle is intentionally not stored as this represents a reasonable architectural tradeoff for a long-running cleanup task.

Learnt from: PeaBrane
PR: ai-dynamo/dynamo#1285
File: lib/llm/src/kv_router/scoring.rs:58-63
Timestamp: 2025-05-30T06:38:09.630Z
Learning: In lib/llm/src/kv_router/scoring.rs, the user prefers to keep the panic behavior when calculating load_avg and variance with empty endpoints rather than adding guards for division by zero. They want the code to fail fast on this error condition.

Learnt from: jthomson04
PR: ai-dynamo/dynamo#1429
File: lib/runtime/src/utils/leader_worker_barrier.rs:69-72
Timestamp: 2025-06-08T03:12:03.985Z
Learning: In the leader-worker barrier implementation in lib/runtime/src/utils/leader_worker_barrier.rs, the `wait_for_key_count` function correctly uses exact equality (`==`) instead of greater-than-or-equal (`>=`) because worker IDs must be unique (enforced by etcd create-only operations), ensuring exactly the expected number of workers can register.

🧬 Code Graph Analysis (1)

lib/llm/src/kv_router.rs (3)

lib/llm/src/kv_router/approx.rs (1)

block_size (286-288)

lib/bindings/python/rust/llm/kv.rs (2)

block_size (419-421)

block_size (480-482)

lib/llm/src/kv_router/indexer.rs (2)

block_size (614-616)

block_size (805-807)

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (4)

GitHub Check: Mirror Repository to GitLab
GitHub Check: Build and Test - vllm
GitHub Check: pre-merge-rust (lib/bindings/python)
GitHub Check: pre-merge-rust (.)

🔇 Additional comments (13)

components/metrics/src/bin/mock_worker.rs (1)

96-96: LGTM!

The cast to u32 correctly aligns with the type change in KVHitRateEvent struct, and is safe given the bounded range (0-100) of the randomly generated value.

components/metrics/src/main.rs (1)

202-202: LGTM!

The cast from u32 to usize is safe and maintains compatibility with the existing update_kv_hit_rate method signature.

lib/llm/src/kv_router/protocols.rs (1)

39-39: LGTM!

The type change to u32 standardizes the representation of block counts across the system and provides sufficient range for practical use cases.

lib/llm/src/kv_router/scheduler.rs (3)

39-39: Structural changes look good

The type changes and new fields for token-level tracking are well-structured and consistent with the PR objectives.

Also applies to: 81-81, 87-89

217-218: Method signature updates are correct

The updates to use potential_blocks_and_tokens and the expanded add_request parameters properly integrate with the new token-level tracking.

Also applies to: 241-246

390-394: Verify the new scheduling formula

The scheduling formula has fundamentally changed from considering cache reuse (request_blocks - cached_blocks) to considering total load (potential_prefill_blocks + potential_blocks). This appears to prioritize load balancing over cache efficiency.

Please confirm this change aligns with the intended scheduling strategy. The new formula:

Weights prefill load (tokens) alongside decode load (blocks)

No longer directly factors in cache hits (overlap_blocks)

May distribute load more evenly but potentially reduce cache reuse benefits

Consider documenting this scheduling strategy change in comments or configuration.

lib/llm/src/kv_router/sequence.rs (7)

80-81: LGTM! Token tracking fields are properly structured.

The addition of prefill_tokens and active_tokens fields enables fine-grained token-level tracking alongside the existing block-level tracking, which aligns with the PR objectives.

Also applies to: 90-92

142-162: Well-implemented token tracking in add_request.

The method correctly calculates new tokens excluding overlap and maintains both token and block accounting. The implementation aligns with the PR objective of tracking in-progress prefill tokens.

173-182: Clean implementation of combined metrics.

The method efficiently reuses existing calculations and provides a unified interface for retrieving both block and token metrics.

201-206: Correct token accounting in free method.

The implementation properly handles both prefill and decode states, with appropriate fallback to 1 token for decode requests.

229-238: Proper state transition from prefill to decode.

The implementation correctly handles the transition when tokens are pushed, updating the active token count to reflect the decode state.

552-588: Well-designed multi-worker token metrics collection.

The implementation efficiently queries all workers in parallel with their specific overlap values and properly aggregates the results. The fallback to 0 for missing overlap scores is reasonable.

622-630: Test coverage properly validates token tracking.

The tests have been appropriately updated to verify both token and block accounting with different overlap scenarios.

Also applies to: 659-659, 675-683

lib/llm/src/kv_router.rs

lib/llm/src/kv_router/sequence.rs

commit e330d96 Author: Yan Ru Pei <yanrpei@gmail.com> Date: Fri Jul 18 13:40:54 2025 -0700 feat: enable / disable chunked prefill for mockers (#2015) Signed-off-by: Yan Ru Pei <yanrpei@gmail.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> commit 353146e Author: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Date: Fri Jul 18 13:33:36 2025 -0700 feat: add vLLM v1 multi-modal example. Add llama4 Maverick example (#1990) Signed-off-by: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Co-authored-by: krishung5 <krish@nvidia.com> commit 1f07dab Author: Jacky <18255193+kthui@users.noreply.github.com> Date: Fri Jul 18 13:04:20 2025 -0700 feat: Add migration to LLM requests (#1930) commit 5f17918 Author: Tanmay Verma <tanmayv@nvidia.com> Date: Fri Jul 18 12:59:34 2025 -0700 refactor: Migrate to new UX2 for python launch (#2003) commit fc12436 Author: Graham King <grahamk@nvidia.com> Date: Fri Jul 18 14:52:57 2025 -0400 feat(frontend): router-mode settings (#2001) commit dc75cf1 Author: ptarasiewiczNV <104908264+ptarasiewiczNV@users.noreply.github.com> Date: Fri Jul 18 18:47:28 2025 +0200 chore: Move NIXL repo clone to Dockerfiles (#2009) commit f6f392c Author: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> Date: Thu Jul 17 18:44:17 2025 -0700 Remove link to the fix for disagg + eagle3 for TRT-LLM example (#2006) Signed-off-by: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> commit cc90ca6 Author: atchernych <atchernych@nvidia.com> Date: Thu Jul 17 18:34:40 2025 -0700 feat: Create a convenience script to uninstall Dynamo Deploy CRDs (#1933) commit 267b422 Author: Greg Clark <grclark@nvidia.com> Date: Thu Jul 17 20:44:21 2025 -0400 chore: loosed python requirement versions (#1998) Signed-off-by: Greg Clark <grclark@nvidia.com> commit b8474e5 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Thu Jul 17 16:35:05 2025 -0700 chore: update cmake and gap installation and sgl in wideep container (#1991) commit 157a3b0 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 15:38:12 2025 -0700 fix: incorrect helm upgrade command (#2000) commit 0dfca2c Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 15:33:33 2025 -0700 ci: Update trtllm gitlab triggers for new components directory and test script (#1992) commit f3fb09e Author: Kris Hung <krish@nvidia.com> Date: Thu Jul 17 14:59:59 2025 -0700 fix: Fix syntax for tokio-console (#1997) commit dacffb8 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 14:57:10 2025 -0700 fix: use non-dev golang image for operator (#1993) commit 2b29a0a Author: zaristei <zaristei@berkeley.edu> Date: Thu Jul 17 13:10:42 2025 -0700 fix: Working Arm Build Dockerfile for Vllm_v1 (#1844) commit 2430d89 Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 12:57:46 2025 -0700 test: Add trtllm kv router tests (#1988) commit 1eadc01 Author: Graham King <grahamk@nvidia.com> Date: Thu Jul 17 15:07:41 2025 -0400 feat(runtime): Support tokio-console (#1986) commit b62e633 Author: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Date: Thu Jul 17 11:16:28 2025 -0700 feat: support separate chat_template.jinja file (#1853) commit 8ae3719 Author: Hongkuan Zhou <tedzhouhk@gmail.com> Date: Thu Jul 17 11:12:35 2025 -0700 chore: add some details to dynamo deploy quickstart and fix deploy.sh (#1978) Signed-off-by: Hongkuan Zhou <tedzhouhk@gmail.com> Co-authored-by: julienmancuso <161955438+julienmancuso@users.noreply.github.com> commit 08891ff Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 10:57:42 2025 -0700 fix: Update trtllm tests to use new scripts instead of dynamo serve (#1979) commit 49b7a0d Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Thu Jul 17 08:35:04 2025 -0600 feat: record + analyze logprobs (#1957) commit 6d2be14 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 00:17:58 2025 -0700 refactor: replace vllm with vllm_v1 container (#1953) Co-authored-by: alec-flowers <aflowers@nvidia.com> commit 4d2a31a Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Wed Jul 16 18:04:09 2025 -0700 chore: add port reservation to utils (#1980) commit 1e3e4a0 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Wed Jul 16 15:54:04 2025 -0700 fix: port race condition through deterministic ports (#1937) commit 4ad281f Author: Tanmay Verma <tanmayv@nvidia.com> Date: Wed Jul 16 14:33:51 2025 -0700 refactor: Move TRTLLM example to the component/backends (#1976) commit 57d24a1 Author: Misha Chornyi <99709299+mc-nv@users.noreply.github.com> Date: Wed Jul 16 14:10:24 2025 -0700 build: Removing shell configuration violations. It's bad practice to hardcod… (#1973) commit 182d3b5 Author: Graham King <grahamk@nvidia.com> Date: Wed Jul 16 16:12:40 2025 -0400 chore(bindings): Remove mistralrs / llama.cpp (#1970) commit def6eaa Author: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Date: Wed Jul 16 15:50:23 2025 -0400 feat: attributions for debian deps of sglang, trtllm, vllm runtime containers (#1971) commit f31732a Author: Yan Ru Pei <yanrpei@gmail.com> Date: Wed Jul 16 11:22:15 2025 -0700 feat: integrate mocker with dynamo-run and python cli (#1927) commit aba6099 Author: Graham King <grahamk@nvidia.com> Date: Wed Jul 16 12:26:32 2025 -0400 perf(router): Remove lock from router hot path (#1963) commit b212103 Author: Hongkuan Zhou <tedzhouhk@gmail.com> Date: Wed Jul 16 08:55:33 2025 -0700 docs: add notes in docs to deprecate local connector (#1959) commit 7b325ee Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 18:52:00 2025 -0700 fix: vllm router examples (#1942) commit a50be1a Author: hhzhang16 <54051230+hhzhang16@users.noreply.github.com> Date: Tue Jul 15 17:58:01 2025 -0700 feat: update CODEOWNERS (#1926) commit e260fdf Author: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Date: Tue Jul 15 18:49:21 2025 -0400 feat: add bitnami helm chart attribution (#1943) Signed-off-by: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> commit 1c03404 Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 14:26:24 2025 -0700 fix: update inference gateway deployment instructions (#1940) commit 5ca570f Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:54:03 2025 -0400 chore: Rename dynamo.ingress to dynamo.frontend (#1944) commit 7b9182f Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:33:07 2025 -0400 chore: Move examples/cli to lib/bindings/examples/cli (#1952) commit 40d40dd Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:02:19 2025 -0400 chore(multi-modal): Rename frontend.py to web.py (#1951) commit a9e0891 Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Tue Jul 15 12:30:30 2025 -0600 feat: adding http clients and recorded response stream (#1919) commit 4128d58 Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 10:30:47 2025 -0700 feat: allow helm upgrade using deploy script (#1936) commit 4da078b Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 12:57:38 2025 -0400 fix: Remove OpenSSL dependency, use Rust TLS (#1945) commit fc004d4 Author: jthomson04 <jwillthomson19@gmail.com> Date: Tue Jul 15 08:45:42 2025 -0700 fix: Fix TRT-LLM container build when using a custom pip wheel (#1825) commit 3c6fc6f Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 22:35:20 2025 -0700 chore: fix typo (#1938) commit de7fe38 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Mon Jul 14 21:47:12 2025 -0700 feat: add vllm e2e integration tests (#1935) commit 860f3f7 Author: Keiven C <213854356+keivenchang@users.noreply.github.com> Date: Mon Jul 14 21:44:19 2025 -0700 chore: metrics endpoint variables renamed from HTTP_SERVER->SYSTEM (#1934) Co-authored-by: Keiven Chang <keivenchang@users.noreply.github.com> commit fc402a3 Author: Biswa Panda <biswa.panda@gmail.com> Date: Mon Jul 14 21:21:20 2025 -0700 feat: configurable namespace for vllm v1 example (#1909) commit df40d2c Author: ZichengMa <zichengma1225@gmail.com> Date: Mon Jul 14 21:11:29 2025 -0700 docs: fix typo and add mount-workspace to vllm doc (#1931) Signed-off-by: ZichengMa <zichengma1225@gmail.com> Co-authored-by: Alec <35311602+alec-flowers@users.noreply.github.com> commit 901715b Author: Tanmay Verma <tanmayv@nvidia.com> Date: Mon Jul 14 20:14:51 2025 -0700 refactor: Refactor the TRTLLM examples remove dynamo SDK (#1884) commit 5bf23d5 Author: hhzhang16 <54051230+hhzhang16@users.noreply.github.com> Date: Mon Jul 14 18:29:19 2025 -0700 feat: update DynamoGraphDeployments for vllm_v1 (#1890) Co-authored-by: mohammedabdulwahhab <furkhan324@berkeley.edu> commit 9e76590 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 17:29:56 2025 -0700 docs: organize sglang readme (#1910) commit ef59ac8 Author: KrishnanPrash <140860868+KrishnanPrash@users.noreply.github.com> Date: Mon Jul 14 16:16:44 2025 -0700 docs: TRTLLM Example of Llama4+Eagle3 (Speculative Decoding) (#1828) Signed-off-by: KrishnanPrash <140860868+KrishnanPrash@users.noreply.github.com> Co-authored-by: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> commit 053041e Author: Jorge António <matroid@outlook.com> Date: Tue Jul 15 00:06:38 2025 +0100 fix: resolve incorrect finish reason propagation (#1857) commit 3733f58 Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 19:04:22 2025 -0400 feat(backends): Python llama.cpp engine (#1925) commit 6a1350c Author: Tushar Sharma <tusharma@nvidia.com> Date: Mon Jul 14 14:56:36 2025 -0700 build: minor improvements to sglang dockerfile (#1917) commit e2a619b Author: Neelay Shah <neelays@nvidia.com> Date: Mon Jul 14 14:52:53 2025 -0700 fix: remove environment variable passing (#1911) Signed-off-by: Neelay Shah <neelays@nvidia.com> Co-authored-by: Neelay Shah <neelays@a4u8g-0057.ipp2u2.colossus.nvidia.com> commit 3d17a49 Author: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> Date: Mon Jul 14 14:41:56 2025 -0700 refactor: remove dynamo build (#1778) Signed-off-by: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> commit 3e0cb07 Author: Anant Sharma <anants@nvidia.com> Date: Mon Jul 14 15:43:48 2025 -0400 fix: copy attributions and license to trtllm runtime container (#1916) commit fc36bf5 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 12:31:49 2025 -0700 feat: receive kvmetrics from sglang scheduler (#1789) Co-authored-by: zixuanzhang226 <zixuanzhang@bytedance.com> commit df91fce Author: Yan Ru Pei <yanrpei@gmail.com> Date: Mon Jul 14 12:24:04 2025 -0700 feat: prefill aware routing (#1895) commit ad8ad66 Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 15:20:35 2025 -0400 feat: Shrink the ai-dynamo wheel by 35 MiB (#1918) Remove http and llmctl binaries. They have been unused for a while. commit 480b41d Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 15:06:45 2025 -0400 feat: Python frontend / ingress node (#1912)

commit cb6de94 Author: ptarasiewiczNV <104908264+ptarasiewiczNV@users.noreply.github.com> Date: Sun Jul 20 22:34:50 2025 +0200 chore: Install vLLM and WideEP kernels in vLLM runtime container (#2010) Signed-off-by: Alec <35311602+alec-flowers@users.noreply.github.com> Co-authored-by: Alec <35311602+alec-flowers@users.noreply.github.com> Co-authored-by: alec-flowers <aflowers@nvidia.com> commit fe63c17 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Fri Jul 18 17:45:08 2025 -0700 fix: Revert "feat: add vLLM v1 multi-modal example. Add llama4 Maverick ex… (#2017) commit bf1998f Author: jthomson04 <jwillthomson19@gmail.com> Date: Fri Jul 18 17:23:50 2025 -0700 fix: Don't detokenize twice in TRT-LLM examples (#1955) commit 343a481 Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Fri Jul 18 16:22:43 2025 -0600 feat: http disconnects (#2014) commit e330d96 Author: Yan Ru Pei <yanrpei@gmail.com> Date: Fri Jul 18 13:40:54 2025 -0700 feat: enable / disable chunked prefill for mockers (#2015) Signed-off-by: Yan Ru Pei <yanrpei@gmail.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> commit 353146e Author: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Date: Fri Jul 18 13:33:36 2025 -0700 feat: add vLLM v1 multi-modal example. Add llama4 Maverick example (#1990) Signed-off-by: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Co-authored-by: krishung5 <krish@nvidia.com> commit 1f07dab Author: Jacky <18255193+kthui@users.noreply.github.com> Date: Fri Jul 18 13:04:20 2025 -0700 feat: Add migration to LLM requests (#1930) commit 5f17918 Author: Tanmay Verma <tanmayv@nvidia.com> Date: Fri Jul 18 12:59:34 2025 -0700 refactor: Migrate to new UX2 for python launch (#2003) commit fc12436 Author: Graham King <grahamk@nvidia.com> Date: Fri Jul 18 14:52:57 2025 -0400 feat(frontend): router-mode settings (#2001) commit dc75cf1 Author: ptarasiewiczNV <104908264+ptarasiewiczNV@users.noreply.github.com> Date: Fri Jul 18 18:47:28 2025 +0200 chore: Move NIXL repo clone to Dockerfiles (#2009) commit f6f392c Author: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> Date: Thu Jul 17 18:44:17 2025 -0700 Remove link to the fix for disagg + eagle3 for TRT-LLM example (#2006) Signed-off-by: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> commit cc90ca6 Author: atchernych <atchernych@nvidia.com> Date: Thu Jul 17 18:34:40 2025 -0700 feat: Create a convenience script to uninstall Dynamo Deploy CRDs (#1933) commit 267b422 Author: Greg Clark <grclark@nvidia.com> Date: Thu Jul 17 20:44:21 2025 -0400 chore: loosed python requirement versions (#1998) Signed-off-by: Greg Clark <grclark@nvidia.com> commit b8474e5 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Thu Jul 17 16:35:05 2025 -0700 chore: update cmake and gap installation and sgl in wideep container (#1991) commit 157a3b0 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 15:38:12 2025 -0700 fix: incorrect helm upgrade command (#2000) commit 0dfca2c Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 15:33:33 2025 -0700 ci: Update trtllm gitlab triggers for new components directory and test script (#1992) commit f3fb09e Author: Kris Hung <krish@nvidia.com> Date: Thu Jul 17 14:59:59 2025 -0700 fix: Fix syntax for tokio-console (#1997) commit dacffb8 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 14:57:10 2025 -0700 fix: use non-dev golang image for operator (#1993) commit 2b29a0a Author: zaristei <zaristei@berkeley.edu> Date: Thu Jul 17 13:10:42 2025 -0700 fix: Working Arm Build Dockerfile for Vllm_v1 (#1844) commit 2430d89 Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 12:57:46 2025 -0700 test: Add trtllm kv router tests (#1988) commit 1eadc01 Author: Graham King <grahamk@nvidia.com> Date: Thu Jul 17 15:07:41 2025 -0400 feat(runtime): Support tokio-console (#1986) commit b62e633 Author: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Date: Thu Jul 17 11:16:28 2025 -0700 feat: support separate chat_template.jinja file (#1853) commit 8ae3719 Author: Hongkuan Zhou <tedzhouhk@gmail.com> Date: Thu Jul 17 11:12:35 2025 -0700 chore: add some details to dynamo deploy quickstart and fix deploy.sh (#1978) Signed-off-by: Hongkuan Zhou <tedzhouhk@gmail.com> Co-authored-by: julienmancuso <161955438+julienmancuso@users.noreply.github.com> commit 08891ff Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 10:57:42 2025 -0700 fix: Update trtllm tests to use new scripts instead of dynamo serve (#1979) commit 49b7a0d Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Thu Jul 17 08:35:04 2025 -0600 feat: record + analyze logprobs (#1957) commit 6d2be14 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 00:17:58 2025 -0700 refactor: replace vllm with vllm_v1 container (#1953) Co-authored-by: alec-flowers <aflowers@nvidia.com> commit 4d2a31a Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Wed Jul 16 18:04:09 2025 -0700 chore: add port reservation to utils (#1980) commit 1e3e4a0 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Wed Jul 16 15:54:04 2025 -0700 fix: port race condition through deterministic ports (#1937) commit 4ad281f Author: Tanmay Verma <tanmayv@nvidia.com> Date: Wed Jul 16 14:33:51 2025 -0700 refactor: Move TRTLLM example to the component/backends (#1976) commit 57d24a1 Author: Misha Chornyi <99709299+mc-nv@users.noreply.github.com> Date: Wed Jul 16 14:10:24 2025 -0700 build: Removing shell configuration violations. It's bad practice to hardcod… (#1973) commit 182d3b5 Author: Graham King <grahamk@nvidia.com> Date: Wed Jul 16 16:12:40 2025 -0400 chore(bindings): Remove mistralrs / llama.cpp (#1970) commit def6eaa Author: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Date: Wed Jul 16 15:50:23 2025 -0400 feat: attributions for debian deps of sglang, trtllm, vllm runtime containers (#1971) commit f31732a Author: Yan Ru Pei <yanrpei@gmail.com> Date: Wed Jul 16 11:22:15 2025 -0700 feat: integrate mocker with dynamo-run and python cli (#1927) commit aba6099 Author: Graham King <grahamk@nvidia.com> Date: Wed Jul 16 12:26:32 2025 -0400 perf(router): Remove lock from router hot path (#1963) commit b212103 Author: Hongkuan Zhou <tedzhouhk@gmail.com> Date: Wed Jul 16 08:55:33 2025 -0700 docs: add notes in docs to deprecate local connector (#1959) commit 7b325ee Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 18:52:00 2025 -0700 fix: vllm router examples (#1942) commit a50be1a Author: hhzhang16 <54051230+hhzhang16@users.noreply.github.com> Date: Tue Jul 15 17:58:01 2025 -0700 feat: update CODEOWNERS (#1926) commit e260fdf Author: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Date: Tue Jul 15 18:49:21 2025 -0400 feat: add bitnami helm chart attribution (#1943) Signed-off-by: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> commit 1c03404 Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 14:26:24 2025 -0700 fix: update inference gateway deployment instructions (#1940) commit 5ca570f Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:54:03 2025 -0400 chore: Rename dynamo.ingress to dynamo.frontend (#1944) commit 7b9182f Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:33:07 2025 -0400 chore: Move examples/cli to lib/bindings/examples/cli (#1952) commit 40d40dd Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:02:19 2025 -0400 chore(multi-modal): Rename frontend.py to web.py (#1951) commit a9e0891 Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Tue Jul 15 12:30:30 2025 -0600 feat: adding http clients and recorded response stream (#1919) commit 4128d58 Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 10:30:47 2025 -0700 feat: allow helm upgrade using deploy script (#1936) commit 4da078b Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 12:57:38 2025 -0400 fix: Remove OpenSSL dependency, use Rust TLS (#1945) commit fc004d4 Author: jthomson04 <jwillthomson19@gmail.com> Date: Tue Jul 15 08:45:42 2025 -0700 fix: Fix TRT-LLM container build when using a custom pip wheel (#1825) commit 3c6fc6f Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 22:35:20 2025 -0700 chore: fix typo (#1938) commit de7fe38 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Mon Jul 14 21:47:12 2025 -0700 feat: add vllm e2e integration tests (#1935) commit 860f3f7 Author: Keiven C <213854356+keivenchang@users.noreply.github.com> Date: Mon Jul 14 21:44:19 2025 -0700 chore: metrics endpoint variables renamed from HTTP_SERVER->SYSTEM (#1934) Co-authored-by: Keiven Chang <keivenchang@users.noreply.github.com> commit fc402a3 Author: Biswa Panda <biswa.panda@gmail.com> Date: Mon Jul 14 21:21:20 2025 -0700 feat: configurable namespace for vllm v1 example (#1909) commit df40d2c Author: ZichengMa <zichengma1225@gmail.com> Date: Mon Jul 14 21:11:29 2025 -0700 docs: fix typo and add mount-workspace to vllm doc (#1931) Signed-off-by: ZichengMa <zichengma1225@gmail.com> Co-authored-by: Alec <35311602+alec-flowers@users.noreply.github.com> commit 901715b Author: Tanmay Verma <tanmayv@nvidia.com> Date: Mon Jul 14 20:14:51 2025 -0700 refactor: Refactor the TRTLLM examples remove dynamo SDK (#1884) commit 5bf23d5 Author: hhzhang16 <54051230+hhzhang16@users.noreply.github.com> Date: Mon Jul 14 18:29:19 2025 -0700 feat: update DynamoGraphDeployments for vllm_v1 (#1890) Co-authored-by: mohammedabdulwahhab <furkhan324@berkeley.edu> commit 9e76590 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 17:29:56 2025 -0700 docs: organize sglang readme (#1910) commit ef59ac8 Author: KrishnanPrash <140860868+KrishnanPrash@users.noreply.github.com> Date: Mon Jul 14 16:16:44 2025 -0700 docs: TRTLLM Example of Llama4+Eagle3 (Speculative Decoding) (#1828) Signed-off-by: KrishnanPrash <140860868+KrishnanPrash@users.noreply.github.com> Co-authored-by: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> commit 053041e Author: Jorge António <matroid@outlook.com> Date: Tue Jul 15 00:06:38 2025 +0100 fix: resolve incorrect finish reason propagation (#1857) commit 3733f58 Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 19:04:22 2025 -0400 feat(backends): Python llama.cpp engine (#1925) commit 6a1350c Author: Tushar Sharma <tusharma@nvidia.com> Date: Mon Jul 14 14:56:36 2025 -0700 build: minor improvements to sglang dockerfile (#1917) commit e2a619b Author: Neelay Shah <neelays@nvidia.com> Date: Mon Jul 14 14:52:53 2025 -0700 fix: remove environment variable passing (#1911) Signed-off-by: Neelay Shah <neelays@nvidia.com> Co-authored-by: Neelay Shah <neelays@a4u8g-0057.ipp2u2.colossus.nvidia.com> commit 3d17a49 Author: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> Date: Mon Jul 14 14:41:56 2025 -0700 refactor: remove dynamo build (#1778) Signed-off-by: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> commit 3e0cb07 Author: Anant Sharma <anants@nvidia.com> Date: Mon Jul 14 15:43:48 2025 -0400 fix: copy attributions and license to trtllm runtime container (#1916) commit fc36bf5 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 12:31:49 2025 -0700 feat: receive kvmetrics from sglang scheduler (#1789) Co-authored-by: zixuanzhang226 <zixuanzhang@bytedance.com> commit df91fce Author: Yan Ru Pei <yanrpei@gmail.com> Date: Mon Jul 14 12:24:04 2025 -0700 feat: prefill aware routing (#1895) commit ad8ad66 Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 15:20:35 2025 -0400 feat: Shrink the ai-dynamo wheel by 35 MiB (#1918) Remove http and llmctl binaries. They have been unused for a while. commit 480b41d Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 15:06:45 2025 -0400 feat: Python frontend / ingress node (#1912)

commit d4b5414 Author: atchernych <atchernych@nvidia.com> Date: Mon Jul 21 13:10:24 2025 -0700 fix: mypy error (#2029) commit 79337c7 Author: Ryan McCormick <rmccormick@nvidia.com> Date: Mon Jul 21 12:12:16 2025 -0700 build: support custom TRTLLM build for commits not on main branch (#2021) commit 95dd942 Author: atchernych <atchernych@nvidia.com> Date: Mon Jul 21 12:09:33 2025 -0700 docs: Post-Merge cleanup of the deploy documentation (#1922) commit cb6de94 Author: ptarasiewiczNV <104908264+ptarasiewiczNV@users.noreply.github.com> Date: Sun Jul 20 22:34:50 2025 +0200 chore: Install vLLM and WideEP kernels in vLLM runtime container (#2010) Signed-off-by: Alec <35311602+alec-flowers@users.noreply.github.com> Co-authored-by: Alec <35311602+alec-flowers@users.noreply.github.com> Co-authored-by: alec-flowers <aflowers@nvidia.com> commit fe63c17 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Fri Jul 18 17:45:08 2025 -0700 fix: Revert "feat: add vLLM v1 multi-modal example. Add llama4 Maverick ex… (#2017) commit bf1998f Author: jthomson04 <jwillthomson19@gmail.com> Date: Fri Jul 18 17:23:50 2025 -0700 fix: Don't detokenize twice in TRT-LLM examples (#1955) commit 343a481 Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Fri Jul 18 16:22:43 2025 -0600 feat: http disconnects (#2014) commit e330d96 Author: Yan Ru Pei <yanrpei@gmail.com> Date: Fri Jul 18 13:40:54 2025 -0700 feat: enable / disable chunked prefill for mockers (#2015) Signed-off-by: Yan Ru Pei <yanrpei@gmail.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> commit 353146e Author: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Date: Fri Jul 18 13:33:36 2025 -0700 feat: add vLLM v1 multi-modal example. Add llama4 Maverick example (#1990) Signed-off-by: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Co-authored-by: krishung5 <krish@nvidia.com> commit 1f07dab Author: Jacky <18255193+kthui@users.noreply.github.com> Date: Fri Jul 18 13:04:20 2025 -0700 feat: Add migration to LLM requests (#1930) commit 5f17918 Author: Tanmay Verma <tanmayv@nvidia.com> Date: Fri Jul 18 12:59:34 2025 -0700 refactor: Migrate to new UX2 for python launch (#2003) commit fc12436 Author: Graham King <grahamk@nvidia.com> Date: Fri Jul 18 14:52:57 2025 -0400 feat(frontend): router-mode settings (#2001) commit dc75cf1 Author: ptarasiewiczNV <104908264+ptarasiewiczNV@users.noreply.github.com> Date: Fri Jul 18 18:47:28 2025 +0200 chore: Move NIXL repo clone to Dockerfiles (#2009) commit f6f392c Author: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> Date: Thu Jul 17 18:44:17 2025 -0700 Remove link to the fix for disagg + eagle3 for TRT-LLM example (#2006) Signed-off-by: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> commit cc90ca6 Author: atchernych <atchernych@nvidia.com> Date: Thu Jul 17 18:34:40 2025 -0700 feat: Create a convenience script to uninstall Dynamo Deploy CRDs (#1933) commit 267b422 Author: Greg Clark <grclark@nvidia.com> Date: Thu Jul 17 20:44:21 2025 -0400 chore: loosed python requirement versions (#1998) Signed-off-by: Greg Clark <grclark@nvidia.com> commit b8474e5 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Thu Jul 17 16:35:05 2025 -0700 chore: update cmake and gap installation and sgl in wideep container (#1991) commit 157a3b0 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 15:38:12 2025 -0700 fix: incorrect helm upgrade command (#2000) commit 0dfca2c Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 15:33:33 2025 -0700 ci: Update trtllm gitlab triggers for new components directory and test script (#1992) commit f3fb09e Author: Kris Hung <krish@nvidia.com> Date: Thu Jul 17 14:59:59 2025 -0700 fix: Fix syntax for tokio-console (#1997) commit dacffb8 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 14:57:10 2025 -0700 fix: use non-dev golang image for operator (#1993) commit 2b29a0a Author: zaristei <zaristei@berkeley.edu> Date: Thu Jul 17 13:10:42 2025 -0700 fix: Working Arm Build Dockerfile for Vllm_v1 (#1844) commit 2430d89 Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 12:57:46 2025 -0700 test: Add trtllm kv router tests (#1988) commit 1eadc01 Author: Graham King <grahamk@nvidia.com> Date: Thu Jul 17 15:07:41 2025 -0400 feat(runtime): Support tokio-console (#1986) commit b62e633 Author: GuanLuo <41310872+GuanLuo@users.noreply.github.com> Date: Thu Jul 17 11:16:28 2025 -0700 feat: support separate chat_template.jinja file (#1853) commit 8ae3719 Author: Hongkuan Zhou <tedzhouhk@gmail.com> Date: Thu Jul 17 11:12:35 2025 -0700 chore: add some details to dynamo deploy quickstart and fix deploy.sh (#1978) Signed-off-by: Hongkuan Zhou <tedzhouhk@gmail.com> Co-authored-by: julienmancuso <161955438+julienmancuso@users.noreply.github.com> commit 08891ff Author: Ryan McCormick <rmccormick@nvidia.com> Date: Thu Jul 17 10:57:42 2025 -0700 fix: Update trtllm tests to use new scripts instead of dynamo serve (#1979) commit 49b7a0d Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Thu Jul 17 08:35:04 2025 -0600 feat: record + analyze logprobs (#1957) commit 6d2be14 Author: Biswa Panda <biswa.panda@gmail.com> Date: Thu Jul 17 00:17:58 2025 -0700 refactor: replace vllm with vllm_v1 container (#1953) Co-authored-by: alec-flowers <aflowers@nvidia.com> commit 4d2a31a Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Wed Jul 16 18:04:09 2025 -0700 chore: add port reservation to utils (#1980) commit 1e3e4a0 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Wed Jul 16 15:54:04 2025 -0700 fix: port race condition through deterministic ports (#1937) commit 4ad281f Author: Tanmay Verma <tanmayv@nvidia.com> Date: Wed Jul 16 14:33:51 2025 -0700 refactor: Move TRTLLM example to the component/backends (#1976) commit 57d24a1 Author: Misha Chornyi <99709299+mc-nv@users.noreply.github.com> Date: Wed Jul 16 14:10:24 2025 -0700 build: Removing shell configuration violations. It's bad practice to hardcod… (#1973) commit 182d3b5 Author: Graham King <grahamk@nvidia.com> Date: Wed Jul 16 16:12:40 2025 -0400 chore(bindings): Remove mistralrs / llama.cpp (#1970) commit def6eaa Author: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Date: Wed Jul 16 15:50:23 2025 -0400 feat: attributions for debian deps of sglang, trtllm, vllm runtime containers (#1971) commit f31732a Author: Yan Ru Pei <yanrpei@gmail.com> Date: Wed Jul 16 11:22:15 2025 -0700 feat: integrate mocker with dynamo-run and python cli (#1927) commit aba6099 Author: Graham King <grahamk@nvidia.com> Date: Wed Jul 16 12:26:32 2025 -0400 perf(router): Remove lock from router hot path (#1963) commit b212103 Author: Hongkuan Zhou <tedzhouhk@gmail.com> Date: Wed Jul 16 08:55:33 2025 -0700 docs: add notes in docs to deprecate local connector (#1959) commit 7b325ee Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 18:52:00 2025 -0700 fix: vllm router examples (#1942) commit a50be1a Author: hhzhang16 <54051230+hhzhang16@users.noreply.github.com> Date: Tue Jul 15 17:58:01 2025 -0700 feat: update CODEOWNERS (#1926) commit e260fdf Author: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Date: Tue Jul 15 18:49:21 2025 -0400 feat: add bitnami helm chart attribution (#1943) Signed-off-by: Harrison Saturley-Hall <454891+saturley-hall@users.noreply.github.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> commit 1c03404 Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 14:26:24 2025 -0700 fix: update inference gateway deployment instructions (#1940) commit 5ca570f Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:54:03 2025 -0400 chore: Rename dynamo.ingress to dynamo.frontend (#1944) commit 7b9182f Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:33:07 2025 -0400 chore: Move examples/cli to lib/bindings/examples/cli (#1952) commit 40d40dd Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 16:02:19 2025 -0400 chore(multi-modal): Rename frontend.py to web.py (#1951) commit a9e0891 Author: Ryan Olson <ryanolson@users.noreply.github.com> Date: Tue Jul 15 12:30:30 2025 -0600 feat: adding http clients and recorded response stream (#1919) commit 4128d58 Author: Biswa Panda <biswa.panda@gmail.com> Date: Tue Jul 15 10:30:47 2025 -0700 feat: allow helm upgrade using deploy script (#1936) commit 4da078b Author: Graham King <grahamk@nvidia.com> Date: Tue Jul 15 12:57:38 2025 -0400 fix: Remove OpenSSL dependency, use Rust TLS (#1945) commit fc004d4 Author: jthomson04 <jwillthomson19@gmail.com> Date: Tue Jul 15 08:45:42 2025 -0700 fix: Fix TRT-LLM container build when using a custom pip wheel (#1825) commit 3c6fc6f Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 22:35:20 2025 -0700 chore: fix typo (#1938) commit de7fe38 Author: Alec <35311602+alec-flowers@users.noreply.github.com> Date: Mon Jul 14 21:47:12 2025 -0700 feat: add vllm e2e integration tests (#1935) commit 860f3f7 Author: Keiven C <213854356+keivenchang@users.noreply.github.com> Date: Mon Jul 14 21:44:19 2025 -0700 chore: metrics endpoint variables renamed from HTTP_SERVER->SYSTEM (#1934) Co-authored-by: Keiven Chang <keivenchang@users.noreply.github.com> commit fc402a3 Author: Biswa Panda <biswa.panda@gmail.com> Date: Mon Jul 14 21:21:20 2025 -0700 feat: configurable namespace for vllm v1 example (#1909) commit df40d2c Author: ZichengMa <zichengma1225@gmail.com> Date: Mon Jul 14 21:11:29 2025 -0700 docs: fix typo and add mount-workspace to vllm doc (#1931) Signed-off-by: ZichengMa <zichengma1225@gmail.com> Co-authored-by: Alec <35311602+alec-flowers@users.noreply.github.com> commit 901715b Author: Tanmay Verma <tanmayv@nvidia.com> Date: Mon Jul 14 20:14:51 2025 -0700 refactor: Refactor the TRTLLM examples remove dynamo SDK (#1884) commit 5bf23d5 Author: hhzhang16 <54051230+hhzhang16@users.noreply.github.com> Date: Mon Jul 14 18:29:19 2025 -0700 feat: update DynamoGraphDeployments for vllm_v1 (#1890) Co-authored-by: mohammedabdulwahhab <furkhan324@berkeley.edu> commit 9e76590 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 17:29:56 2025 -0700 docs: organize sglang readme (#1910) commit ef59ac8 Author: KrishnanPrash <140860868+KrishnanPrash@users.noreply.github.com> Date: Mon Jul 14 16:16:44 2025 -0700 docs: TRTLLM Example of Llama4+Eagle3 (Speculative Decoding) (#1828) Signed-off-by: KrishnanPrash <140860868+KrishnanPrash@users.noreply.github.com> Co-authored-by: Iman Tabrizian <10105175+Tabrizian@users.noreply.github.com> commit 053041e Author: Jorge António <matroid@outlook.com> Date: Tue Jul 15 00:06:38 2025 +0100 fix: resolve incorrect finish reason propagation (#1857) commit 3733f58 Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 19:04:22 2025 -0400 feat(backends): Python llama.cpp engine (#1925) commit 6a1350c Author: Tushar Sharma <tusharma@nvidia.com> Date: Mon Jul 14 14:56:36 2025 -0700 build: minor improvements to sglang dockerfile (#1917) commit e2a619b Author: Neelay Shah <neelays@nvidia.com> Date: Mon Jul 14 14:52:53 2025 -0700 fix: remove environment variable passing (#1911) Signed-off-by: Neelay Shah <neelays@nvidia.com> Co-authored-by: Neelay Shah <neelays@a4u8g-0057.ipp2u2.colossus.nvidia.com> commit 3d17a49 Author: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> Date: Mon Jul 14 14:41:56 2025 -0700 refactor: remove dynamo build (#1778) Signed-off-by: Schwinn Saereesitthipitak <17022745+galletas1712@users.noreply.github.com> commit 3e0cb07 Author: Anant Sharma <anants@nvidia.com> Date: Mon Jul 14 15:43:48 2025 -0400 fix: copy attributions and license to trtllm runtime container (#1916) commit fc36bf5 Author: ishandhanani <82981111+ishandhanani@users.noreply.github.com> Date: Mon Jul 14 12:31:49 2025 -0700 feat: receive kvmetrics from sglang scheduler (#1789) Co-authored-by: zixuanzhang226 <zixuanzhang@bytedance.com> commit df91fce Author: Yan Ru Pei <yanrpei@gmail.com> Date: Mon Jul 14 12:24:04 2025 -0700 feat: prefill aware routing (#1895) commit ad8ad66 Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 15:20:35 2025 -0400 feat: Shrink the ai-dynamo wheel by 35 MiB (#1918) Remove http and llmctl binaries. They have been unused for a while. commit 480b41d Author: Graham King <grahamk@nvidia.com> Date: Mon Jul 14 15:06:45 2025 -0400 feat: Python frontend / ingress node (#1912)

PeaBrane added 2 commits July 11, 2025 12:42

router queue

0d6bd47

prefill aware

891adaf

pull-request-size bot added the size/L label Jul 11, 2025

copy-pr-bot bot temporarily deployed to GITLAB July 11, 2025 20:54 Inactive

github-actions bot added the feat label Jul 11, 2025

copy-pr-bot bot temporarily deployed to GITLAB July 11, 2025 20:54 Inactive

assuming isl no 0 for prefill tokens

dbf4e91

copy-pr-bot bot temporarily deployed to GITLAB July 11, 2025 21:08 Inactive

assuming isl for decoding blocks also

366c175

copy-pr-bot bot temporarily deployed to GITLAB July 11, 2025 21:17 Inactive

PeaBrane marked this pull request as ready for review July 11, 2025 21:20

PeaBrane requested review from alec-flowers, rmccorm4 and tedzhouhk July 11, 2025 21:21

copy-pr-bot bot temporarily deployed to GITLAB July 11, 2025 21:23 Inactive

coderabbitai bot reviewed Jul 11, 2025

View reviewed changes

lib/llm/src/kv_router.rs Outdated Show resolved Hide resolved

lib/llm/src/kv_router/sequence.rs Outdated Show resolved Hide resolved

rabbit

7bc502f

copy-pr-bot bot temporarily deployed to GITLAB July 11, 2025 21:30 Inactive

copy-pr-bot bot temporarily deployed to GITLAB July 11, 2025 21:31 Inactive

Merge branch 'main' into rupei/router-queue

a6d35a4

copy-pr-bot bot temporarily deployed to GITLAB July 12, 2025 07:21 Inactive

PeaBrane self-assigned this Jul 13, 2025

router temperature default revert back to 0

e5fbff4

copy-pr-bot bot temporarily deployed to GITLAB July 13, 2025 20:58 Inactive

copy-pr-bot bot temporarily deployed to GITLAB July 13, 2025 20:59 Inactive

alec-flowers approved these changes Jul 14, 2025

View reviewed changes

Merge branch 'main' into rupei/router-queue

403e318

copy-pr-bot bot temporarily deployed to GITLAB July 14, 2025 18:44 Inactive

copy-pr-bot bot temporarily deployed to GITLAB July 14, 2025 18:46 Inactive

PeaBrane enabled auto-merge (squash) July 14, 2025 18:46

PeaBrane merged commit df91fce into main Jul 14, 2025
13 of 17 checks passed

PeaBrane deleted the rupei/router-queue branch July 14, 2025 19:24

ZichengMa pushed a commit that referenced this pull request Jul 17, 2025

feat: prefill aware routing (#1895)

644a68d

This was referenced Jul 30, 2025

feat: reduce / revert routing overheads, do not consider output tokens #2182

Merged

feat: Router replicas with state-sharing #2264

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: prefill aware routing #1895

feat: prefill aware routing #1895

Uh oh!

PeaBrane commented Jul 11, 2025 •

edited

Loading

Uh oh!

coderabbitai bot commented Jul 11, 2025

Chat

Support

CodeRabbit Commands (Invoked using PR comments)

Other keywords and placeholders

Documentation and Community

Uh oh!

coderabbitai bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

feat: prefill aware routing #1895

feat: prefill aware routing #1895

Uh oh!

Conversation

PeaBrane commented Jul 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Overview:

Results:

Summary by CodeRabbit

Uh oh!

coderabbitai bot commented Jul 11, 2025

Walkthrough

Changes

Sequence Diagram(s)

Possibly related PRs

Poem

Chat

Support

CodeRabbit Commands (Invoked using PR comments)

Other keywords and placeholders

Documentation and Community

Uh oh!

coderabbitai bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

PeaBrane commented Jul 11, 2025 •

edited

Loading