[#10245][feat] AutoDeploy: Support Finegrained FP8 quantization

[bmarimuthu-nv]

Summary by CodeRabbit

Release Notes

• New Features

  • Added fine-grained FP8 quantization support for linear layers, enabling efficient inference on quantized models.
  • Introduced fine-grained FP8 Mixture of Experts (MoE) fusion for optimized kernel execution.
  • Extended tensor-parallel sharding support to handle fine-grained FP8 quantization.
  • Added checkpoint compatibility for fine-grained FP8 model formats.

• Tests

  • Added comprehensive test coverage for fine-grained FP8 quantization and MoE fusion across different hardware architectures.

Description

Test Coverage

PR Checklist

Please review the following before submitting your PR:

• PR description clearly explains what and why. If using CodeRabbit's summary, please make sure it makes sense.

• PR Follows TRT-LLM CODING GUIDELINES to the best of your knowledge.

• Test cases are provided for new code paths (see test instructions)

• Any new dependencies have been scanned for license and vulnerabilities

• CODEOWNERS updated if ownership changes

• Documentation updated as needed

• Update tava architecture diagram if there is a significant design change in PR.

• The reviewers assigned automatically/manually are appropriate for the PR.

• Please check this after reviewing the above items as appropriate for this PR.

GitHub Bot Help

/bot [-h] ['run', 'kill', 'skip', 'reuse-pipeline'] ...

Provide a user friendly way for developers to interact with a Jenkins server.

Run /bot [-h|--help] to print this help message.

See details below for each supported subcommand.

Details

run [--reuse-test (optional)pipeline-id --disable-fail-fast --skip-test --stage-list "A10-PyTorch-1, xxx" --gpu-type "A30, H100_PCIe" --test-backend "pytorch, cpp" --add-multi-gpu-test --only-multi-gpu-test --disable-multi-gpu-test --post-merge --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" --detailed-log --debug(experimental)]

Launch build/test pipelines. All previously running jobs will be killed.

--reuse-test (optional)pipeline-id (OPTIONAL) : Allow the new pipeline to reuse build artifacts and skip successful test stages from a specified pipeline or the last pipeline if no pipeline-id is indicated. If the Git commit ID has changed, this option will be always ignored. The DEFAULT behavior of the bot is to reuse build artifacts and successful test results from the last pipeline.

--disable-reuse-test (OPTIONAL) : Explicitly prevent the pipeline from reusing build artifacts and skipping successful test stages from a previous pipeline. Ensure that all builds and tests are run regardless of previous successes.

--disable-fail-fast (OPTIONAL) : Disable fail fast on build/tests/infra failures.

--skip-test (OPTIONAL) : Skip all test stages, but still run build stages, package stages and sanity check stages. Note: Does NOT update GitHub check status.

--stage-list "A10-PyTorch-1, xxx" (OPTIONAL) : Only run the specified test stages. Examples: "A10-PyTorch-1, xxx". Note: Does NOT update GitHub check status.

--gpu-type "A30, H100_PCIe" (OPTIONAL) : Only run the test stages on the specified GPU types. Examples: "A30, H100_PCIe". Note: Does NOT update GitHub check status.

--test-backend "pytorch, cpp" (OPTIONAL) : Skip test stages which don't match the specified backends. Only support [pytorch, cpp, tensorrt, triton]. Examples: "pytorch, cpp" (does not run test stages with tensorrt or triton backend). Note: Does NOT update GitHub pipeline status.

--only-multi-gpu-test (OPTIONAL) : Only run the multi-GPU tests. Note: Does NOT update GitHub check status.

--disable-multi-gpu-test (OPTIONAL) : Disable the multi-GPU tests. Note: Does NOT update GitHub check status.

--add-multi-gpu-test (OPTIONAL) : Force run the multi-GPU tests in addition to running L0 pre-merge pipeline.

--post-merge (OPTIONAL) : Run the L0 post-merge pipeline instead of the ordinary L0 pre-merge pipeline.

--extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx" (OPTIONAL) : Run the ordinary L0 pre-merge pipeline and specified test stages. Examples: --extra-stage "H100_PCIe-TensorRT-Post-Merge-1, xxx".

--detailed-log (OPTIONAL) : Enable flushing out all logs to the Jenkins console. This will significantly increase the log volume and may slow down the job.

--debug (OPTIONAL) : Experimental feature. Enable access to the CI container for debugging purpose. Note: Specify exactly one stage in the stage-list parameter to access the appropriate container environment. Note: Does NOT update GitHub check status.

For guidance on mapping tests to stage names, see docs/source/reference/ci-overview.md
and the scripts/test_to_stage_mapping.py helper.

kill

kill

Kill all running builds associated with pull request.

skip

skip --comment COMMENT

Skip testing for latest commit on pull request. --comment "Reason for skipping build/test" is required. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

reuse-pipeline

reuse-pipeline

Reuse a previous pipeline to validate current commit. This action will also kill all currently running builds associated with the pull request. IMPORTANT NOTE: This is dangerous since lack of user care and validation can cause top of tree to break.

[bmarimuthu-nv]

/bot run

[tensorrt-cicd]

PR_Github #33625 [ run ] triggered by Bot. Commit: c6999f1

[tensorrt-cicd]

PR_Github #33625 [ run ] completed with state SUCCESS. Commit: c6999f1
/LLM/main/L0_MergeRequest_PR pipeline #25940 completed with status: 'FAILURE'

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

[bmarimuthu-nv]

@CodeRabbit summary

[coderabbitai]

✅ Actions performed

Summary regeneration triggered.

[coderabbitai]

📝 Walkthrough

Walkthrough

This pull request introduces fine-grained FP8 quantization support for linear and MoE layers, including new custom operations, quantization and fusion transforms, configuration entries, and comprehensive test coverage spanning both PyTorch and TRT-LLM execution paths.

Changes

Cohort / File(s)	Summary
Configuration & Transforms Registry tensorrt_llm/_torch/auto_deploy/config/default.yaml	Adds four new transforms to enable fine-grained FP8 quantization and fusion: quantize_finegrained_fp8_linear_from_config, quantize_finegrained_fp8_moe (pattern_matcher), and fuse_finegrained_fp8_linear, fuse_finegrained_fp8_moe (post_load_fusion).
Custom Ops - Linear Quantization tensorrt_llm/_torch/auto_deploy/custom_ops/quantization/torch_quant.py	Implements torch_fake_quant_finegrained_fp8_linear with block-aware FP8 quantization and trtllm_finegrained_fp8_linear for TRT-LLM optimized execution, including fake-trace variants for export.
Custom Ops - MoE Quantization tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py, tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/trtllm_moe.py	Adds torch_quant_finegrained_fp8_moe for per-expert FP8 quantized MLPs and trtllm_quant_finegrained_fp8_moe_fused with dual Hopper/Blackwell execution paths and kernel dispatch logic.
Quantization Transforms tensorrt_llm/_torch/auto_deploy/transform/library/quantization.py, tensorrt_llm/_torch/auto_deploy/transform/library/quantize_moe.py	Introduces FineGrainedFP8LinearQuantization and QuantizeFineGrainedFP8MOE transforms with per-block weight scale handling and HF checkpoint remapping hooks.
Fusion Transforms tensorrt_llm/_torch/auto_deploy/transform/library/fuse_quant.py, tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py	Implements FuseFineGrainedFP8Linear pattern-matching fusion and FuseFineGrainedFP8Moe weight stacking for per-expert FP8 weights with backend kernel integration.
Sharding Support tensorrt_llm/_torch/auto_deploy/transform/library/sharding.py	Adds FineGrainedFP8WeightShardingInfo and FineGrainedFP8EPShardingInfo for tensor-parallel and expert-parallel sharding with per-block scale handling and quantization skip logic.
Utilities & Readers tensorrt_llm/_torch/auto_deploy/models/quant_config_reader.py, tensorrt_llm/_torch/auto_deploy/utils/...	Updates HFQuantConfigReader to support "fp8" quantization method; extends node detection utilities (is_any_moe_op, is_fake_quantized_linear_op) and quantization skip logic with pattern matching.
Test Models & Fixtures tests/unittest/_torch/auto_deploy/_utils_test/_model_test_utils.py, tests/unittest/_torch/auto_deploy/unit/multigpu/transformations/library/test_tp_sharding.py	Introduces FakeFineGrainedFP8Linear and FineGrainedFP8MLP test model implementations with block-wise FP8 quantization and per-block scale management.
Unit Tests - Quantization & Fusion tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_quant*.py, tests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/quantization/test_quant.py	Adds test coverage for fine-grained FP8 linear quantization, fusion verification, and pattern detection; includes TinyFineGrainedFP8Ref reference model and parameterized fusion tests.
Unit Tests - MoE Fusion & Ops tests/unittest/_torch/auto_deploy/unit/singlegpu/transformations/library/test_moe_fusion.py, tests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/moe/test_trtllm_moe.py	Implements BlockSparseTop2MLPFineGrainedFP8 and FineGrainedFP8MoEOpModel test models; adds fusion and correctness tests with Hopper/Blackwell variant coverage and per-block quantization workflows.
Integration Tests tests/integration/defs/common.py, tests/integration/defs/examples/test_llama.py	Updates LLaMA FP8 conversion paths to use finegrained_fp8 checkpoint variant instead of hf_fp8 identifier.

Sequence Diagram

sequenceDiagram
    participant Config as YAML Config
    participant Quantizer as Quantization Transform
    participant GraphOp as Graph Operations
    participant Kernel as FP8 Kernel<br/>(torch/trtllm)
    participant Fuser as Fusion Transform
    participant Sharder as Sharding Transform

    Config->>Quantizer: Load finegrained_fp8_linear transform
    Quantizer->>GraphOp: Replace linear ops with fake_quant_finegrained_fp8_linear<br/>per-block scales
    GraphOp->>Kernel: Execute with FP8 weights + scales
    Kernel-->>GraphOp: FP8 quantized output
    Fuser->>GraphOp: Detect fake_quant patterns
    Fuser->>Kernel: Fuse into trtllm_finegrained_fp8_linear<br/>or torch variant
    GraphOp->>Sharder: Optimized graph (post-fusion)
    Sharder->>Sharder: Shard weights and per-block scales<br/>for TP/EP
    Sharder-->>GraphOp: Sharded quantized graph

Loading

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~60 minutes

Possibly related PRs

• [#10780][feat] AutoDeploy: Support per-expert scales in FP8 and NVFP4 MoE #11322: Overlapping changes to fine-grained FP8 MoE ops, fusion stacking logic, and allow_different_input_scales configuration.

Suggested reviewers

• lucaslie
• nvchenghaoz

🚥 Pre-merge checks | ✅ 1 | ❌ 2

❌ Failed checks (2 warnings)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The PR description is incomplete. It contains only the repository template with checklist items but lacks actual implementation details about what was changed, why it was changed, which tests validate the changes, and how it addresses the referenced issue.	Complete the description by filling in the 'Description' and 'Test Coverage' sections with specific implementation details and test references that validate the fine-grained FP8 quantization feature.
Docstring Coverage	⚠️ Warning	Docstring coverage is 32.35% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (1 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely summarizes the main change: adding support for fine-grained FP8 quantization to AutoDeploy. It is specific, directly related to the changeset, and follows the repository naming convention.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment

Tip

Issue Planner is now in beta. Read the docs and try it out! Share your feedback on Discord.

---

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

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[bmarimuthu-nv]

/bot run

[tensorrt-cicd]

PR_Github #34613 [ run ] triggered by Bot. Commit: e4e971c

[tensorrt-cicd]

PR_Github #34613 [ run ] completed with state SUCCESS. Commit: e4e971c
/LLM/main/L0_MergeRequest_PR pipeline #26710 completed with status: 'FAILURE'

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

[bmarimuthu-nv]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1"

[tensorrt-cicd]

PR_Github #36097 [ run ] triggered by Bot. Commit: 46fabf8

[bmarimuthu-nv]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1"

[tensorrt-cicd]

PR_Github #36110 [ run ] triggered by Bot. Commit: facbb93

[tensorrt-cicd]

PR_Github #36110 [ run ] completed with state SUCCESS. Commit: facbb93
/LLM/main/L0_MergeRequest_PR pipeline #27903 completed with status: 'FAILURE'

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

[bmarimuthu-nv]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1"

[lucaslie]

lgtm but needs an overhaul now that #10985 has merged

[tensorrt-cicd]

PR_Github #36174 [ run ] triggered by Bot. Commit: 7be5017 Link to invocation

[greg-kwasniewski1]

An open question: should we multiply the function and class variants? It's fine, I guess, to have a separate class, pattern matcher, then torch_X_op, + (fake variants), then trtllm_X_op + (alltoall variants).

But since it appears to me that the finegrained_fp8 quantization is:

• only supported for compute capabilities > SM100
• it falls back to "default" FP8 quantization for hopper
• it diverges from "default" FP8 only in the last stage by calling different quantization of hidden states and torch.ops.trtllm.fp8_block_scale_moe_runner instead of torch.ops.trtllm.fused_moe

Would that make sense if the default/finegrained FP8 quantization be just a flag /runtime check inside trtllm_quant_fp8_moe_fused ?

[bmarimuthu-nv]

An open question: should we multiply the function and class variants? It's fine, I guess, to have a separate class, pattern matcher, then torch_X_op, + (fake variants), then trtllm_X_op + (alltoall variants).

But since it appears to me that the finegrained_fp8 quantization is:

• only supported for compute capabilities > SM100
• it falls back to "default" FP8 quantization for hopper
• it diverges from "default" FP8 only in the last stage by calling different quantization of hidden states and torch.ops.trtllm.fp8_block_scale_moe_runner instead of torch.ops.trtllm.fused_moe

Would that make sense if the default/finegrained FP8 quantization be just a flag /runtime check inside trtllm_quant_fp8_moe_fused ?

I kinda agree that combination of functions for handling class/op/pattern-matcher is increasing. I've cleaned up the MoE alltoall code paths to be simpler now.

But overloading the op seems to make it more complex inside the op. Sure, the Hopper path ends up calling the same backend op, but, still it calls it with entirely different arguments.

Fundamentally, the following things change between the default fp8 and this finegrained FP8:

• The scale formats are fundamentally different shapes ([E] vs [E, N/128, K/128])
• config schemas and Input handling differs (pre-quantized FP8 vs BF16 with dynamic quant)
• Blackwell has a dedicated kernel (fp8_block_scale_moe_runner)

[bmarimuthu-nv]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1" --disable-fail-fast

[tensorrt-cicd]

PR_Github #37774 [ run ] triggered by Bot. Commit: 8b4df4e Link to invocation

[tensorrt-cicd]

PR_Github #37774 [ run ] completed with state SUCCESS. Commit: 8b4df4e
/LLM/main/L0_MergeRequest_PR pipeline #29241 completed with status: 'FAILURE'

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

[bmarimuthu-nv]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1" --disable-fail-fast --reuse-test

[tensorrt-cicd]

PR_Github #37822 [ run ] triggered by Bot. Commit: e0ae4ae Link to invocation

[tensorrt-cicd]

PR_Github #37822 [ run ] completed with state SUCCESS. Commit: e0ae4ae
/LLM/main/L0_MergeRequest_PR pipeline #29284 completed with status: 'FAILURE'

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

Link to invocation

[bmarimuthu-nv]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1" --disable-fail-fast --reuse-test

[tensorrt-cicd]

PR_Github #37895 [ run ] triggered by Bot. Commit: e0ae4ae Link to invocation

[tensorrt-cicd]

PR_Github #37895 [ run ] completed with state SUCCESS. Commit: e0ae4ae
/LLM/main/L0_MergeRequest_PR pipeline #29341 completed with status: 'SUCCESS'

Link to invocation

Greg mentioned offline that the all suggestions are not blocking. I have also made changes & simplified the code paths as per his review.

[bmarimuthu-nv]

/bot run --extra-stage "DGX_B200-4_GPUs-AutoDeploy-1, DGX_H100-4_GPUs-AutoDeploy-1" --disable-fail-fast

[tensorrt-cicd]

PR_Github #37926 [ run ] triggered by Bot. Commit: a27d48d Link to invocation

[tensorrt-cicd]

PR_Github #37926 [ run ] completed with state SUCCESS. Commit: a27d48d
/LLM/main/L0_MergeRequest_PR pipeline #29369 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.

Link to invocation