Enable embedding lookup/lora_a logic for chunked backend

[MogicianWu]

Motivation

Per #14177, we want to support csgmv backend for lora_a shrink forward/embedding lookup.

Modifications

1. Added python/sglang/srt/lora/triton_ops/chunked_embedding_lora_a.py to supported chunked embedding lookup in lora_a operation. Launched #chunks of thread blocks and load embedding along rank dimension.
2. Changed test/registered/lora/test_lora_hf_sgl_logprob_diff.py to use csgmv backend by default.
3. Modified test_shrink_basic inside test/manual/lora/test_chunked_sgmv_backend.py to test lora_a embedding lookup logic as well.
4. Removed lora target module limit in python/sglang/srt/server_args.py. (solved https://github.com/sgl-project/sglang/pull/14796/changes)
5. Added additional triton cache key in python/sglang/srt/lora/triton_ops/chunked_sgmv_expand.py, only use (NUM_SLICES, BLOCK_M) as key will cause collision in test_lora_hf_sgl_logprob_diff.py's test case

 [debug] calling _chunked_lora_expand_kernel, cache key:[num_slices=1, BLOCK_M=16] BLOCK_N=64 BLOCK_K=16 OUTPUT_DIM=4096 MAX_RANK=8
 [debug] calling _chunked_lora_expand_kernel, cache key:[num_slices=1, BLOCK_M=16] BLOCK_N=64 BLOCK_K=16 OUTPUT_DIM=32000 MAX_RANK=8

6. Compute additional lora batch info metadata for use cases in python/sglang/srt/layers/logits_processor.py where the input activation to lm_head module was pruned
7. Added guard inside python/sglang/srt/lora/layers.py to detect pruned lm_head shape mistmatch versus precomputed lora batch shape
8. Support multiple calls to lm_head module with varying chunk sizes after lm_head pruning
9. Added a third test to test/registered/lora/test_lora_hf_sgl_logprob_diff.py to test the above test, added debug logs to see first 2 test cases triggered pruning and the third triggered further chunking

test_lora_logprob_comparison_basic
[DEBUG logits_processor] lm_head pruning: pruned_states.shape=torch.Size([5, 4096]), skip_chunking=True (enable=False, chunk_size=2048)
[DEBUG logits_processor] lm_head pruning: pruned_states.shape=torch.Size([11, 4096]), skip_chunking=True (enable=False, chunk_size=2048)

test_lora_logprob_comparison_full
[DEBUG logits_processor] lm_head pruning: pruned_states.shape=torch.Size([5, 4096]), skip_chunking=True (enable=False, chunk_size=2048)
[DEBUG logits_processor] lm_head pruning: pruned_states.shape=torch.Size([11, 4096]), skip_chunking=True (enable=False, chunk_size=2048)
[DEBUG logits_processor] lm_head pruning: pruned_states.shape=torch.Size([6, 4096]), skip_chunking=True (enable=False, chunk_size=2048)
[DEBUG logits_processor] lm_head pruning: pruned_states.shape=torch.Size([6, 4096]), skip_chunking=True (enable=False, chunk_size=2048)
[DEBUG logits_processor] lm_head pruning: pruned_states.shape=torch.Size([10, 4096]), skip_chunking=True (enable=False, chunk_size=2048)

test_lora_logprob_comparison_chunked
[DEBUG logits_processor] lm_head multi-pass: pruned_states.shape=torch.Size([5, 4096]), chunk_size=4
[DEBUG logits_processor] pass 1/2: tokens [0:4]
[DEBUG logits_processor] pass 2/2: tokens [4:5]
[DEBUG logits_processor] lm_head multi-pass: pruned_states.shape=torch.Size([11, 4096]), chunk_size=4
[DEBUG logits_processor] pass 1/3: tokens [0:4]
[DEBUG logits_processor] pass 2/3: tokens [4:8]
[DEBUG logits_processor] pass 3/3: tokens [8:11]
[DEBUG logits_processor] lm_head multi-pass: pruned_states.shape=torch.Size([6, 4096]), chunk_size=4
[DEBUG logits_processor] pass 1/2: tokens [0:4]
[DEBUG logits_processor] pass 2/2: tokens [4:6]
[DEBUG logits_processor] lm_head multi-pass: pruned_states.shape=torch.Size([6, 4096]), chunk_size=4
[DEBUG logits_processor] pass 1/2: tokens [0:4]
[DEBUG logits_processor] pass 2/2: tokens [4:6]
[DEBUG logits_processor] lm_head multi-pass: pruned_states.shape=torch.Size([10, 4096]), chunk_size=4
[DEBUG logits_processor] pass 1/3: tokens [0:4]
[DEBUG logits_processor] pass 2/3: tokens [4:8]
[DEBUG logits_processor] pass 3/3: tokens [8:10]

Accuracy Tests

Ran test/registered/lora/test_lora_hf_sgl_logprob_diff.py test on H100, passed all tests

 ~/tmp/sglang/sglang (lora_a_csgmv)]$ pytest test/registered/lora/test_lora_hf_sgl_logprob_diff.py -s -vv --log-cli-level=DEBUG
================================================================================
Overall Statistics
================================================================================

Logprob Differences:
  Prefill:
    Max of max:   1.553154e-02
    Mean of max:  1.090739e-02
    Mean of mean: 3.126751e-03
  Decode:
    Max of max:   3.064442e-02
    Mean of max:  2.019897e-02
    Mean of mean: 3.742597e-03

Logprob Statistics (threshold: 1e-01):
  Overall logprob: 5/5 PASSED
  Prefill logprob: 5/5
  Decode logprob:  5/5

String Statistics:
  Output strings:  5/5

================================================================================
Test completed successfully!
================================================================================
PASSED

Benchmarking and Profiling

Ran sanity check using gsm8k

(sglang_env) ~/tmp/sglang/sglang (lora_a_csgmv)]$ python3 -m sglang.test.few_shot_gsm8k --num-questions 200
Downloading from https://raw.githubusercontent.com/openai/grade-school-math/master/grade_school_math/data/test.jsonl to /tmp/test.jsonl
/tmp/test.jsonl: 732kB [00:00, 15.1MB/s]                                                                                                                                                                         
100%|██████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 200/200 [00:05<00:00, 35.16it/s]
Accuracy: 0.860
Invalid: 0.000
Latency: 5.731 s
Output throughput: 5001.777 token/s

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.

• /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci

13. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Summary of Changes

Hello @MogicianWu, 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 significantly enhances the capabilities of the chunked SGMV (csgmv) backend for LoRA operations by introducing comprehensive support for embedding lookup and LoRA A logic. Previously, the csgmv backend had limitations regarding embedding layers. This change integrates a new Triton-based chunked embedding lookup mechanism, removes the associated restrictions, and updates the testing suite to reflect and validate these new functionalities, ensuring accurate and efficient LoRA application to embedding layers within the chunked backend.

Highlights

• Chunked Embedding Lookup for LoRA A: Introduced a new Triton kernel (_chunked_embedding_lora_a_kernel) and its Python wrapper (chunked_embedding_lora_a_forward) to enable efficient chunked embedding lookup for LoRA A operations, processing embedding weights along the rank dimension.
• Expanded csgmv Backend Capabilities: Integrated the new chunked embedding lookup into the ChunkedSgmvLoRABackend, allowing the csgmv backend to fully support LoRA A embedding operations, which was previously a limitation.
• Removed LoRA Target Module Restrictions: Eliminated the previous warning and limitation that prevented the csgmv backend from applying LoRA to embed_tokens and lm_head layers, broadening its applicability to all LoRA target modules.
• Improved Triton Kernel Caching: Enhanced the Triton kernel caching mechanism for chunked_sgmv_expand by including OUTPUT_DIM in the cache key, which prevents potential cache collisions and ensures correct kernel selection.
• Comprehensive Testing Updates: Updated existing tests and added new ones, including a reference_embedding_lora_a_shrink function and a dedicated test case in test_chunked_sgmv_backend.py, to validate the correctness of the chunked embedding LoRA A forward pass. The test_lora_hf_sgl_logprob_diff.py test now defaults to using the csgmv backend.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

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

1. 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. ↩

[gemini-code-assist]

Code Review

This pull request is a solid contribution that enables LoRA for embedding layers in the csgmv backend, removing a previous limitation. The implementation includes a new Triton kernel, a corresponding reference implementation for robust testing, and updates to both unit and end-to-end accuracy tests. I've identified a high-severity bug due to a typo in an __init__.py file and a medium-severity performance concern in the new Triton kernel. Addressing these points will make this an excellent addition to the codebase.

[gemini-code-assist]

There is a mismatch between the imported function name and the name added to __all__. You import chunked_embedding_lora_a_forward but export "chunked_embedding_lora_a_fwd". This will cause a NameError on from ... import * because chunked_embedding_lora_a_fwd is not defined.

For consistency with other chunked operations like chunked_sgmv_lora_expand_forward, the name in __all__ should be "chunked_embedding_lora_a_forward".

Suggested change

	"chunked_embedding_lora_a_fwd",
	"chunked_embedding_lora_a_forward",

[gemini-code-assist]

The for c in range(chunk_start, chunk_end): loop processes tokens within a chunk serially. This can be a performance bottleneck, as Triton kernels are most efficient when they can parallelize operations. Other chunked kernels in this project (e.g., _chunked_lora_shrink_kernel) process blocks of tokens in parallel using tl.arange to achieve better performance.

To improve performance, consider refactoring this kernel to process tokens in parallel within each program instance. While this may require more complex 2D pointer logic for loading from the weights tensor, it would align with high-performance Triton practices and the design of other kernels in this backend.

[Fridge003]

/tag-and-rerun-ci

[Fridge003]

Can we come up with some cleaner way to handling the pruned tokens? The current way is a little invasive

[MogicianWu]

/tag-and-rerun-ci

[yushengsu-thu]

This PR also resolves the issue in #18649.