[None][feat] AutoDeploy: Add FP8 MOE for Nemotron #8599
Conversation
Signed-off-by: Chenghao Zhang <[email protected]>
📝 WalkthroughWalkthroughThis PR introduces FP8 quantized Mixture of Experts (MoE) support to TensorRT LLM's autodeploy system. Changes include enabling a new Changes
Sequence Diagram(s)sequenceDiagram
participant User as User Code
participant Transform as FuseFP8Moe Transform
participant WeightStack as _stack_fp8_moe_weights
participant TritonOp as triton_quant_fp8_moe
participant Kernel as FP8 Kernel
User->>Transform: Process graph
Transform->>WeightStack: Extract FP8 MoE nodes
WeightStack->>WeightStack: Stack per-expert w1, w2, w3 weights
WeightStack->>WeightStack: Stack input/weight scales
WeightStack->>Transform: Register stacked parameters
Transform->>TritonOp: Replace node with triton_quant_fp8_moe call
User->>TritonOp: Forward pass
TritonOp->>Kernel: Invoke FP8 kernel with stacked weights & scales
Kernel->>Kernel: FP8 load & scale<br/>Per-block routing<br/>Scaled accumulate
Kernel-->>TritonOp: Output tensor
TritonOp-->>User: Return result
Estimated code review effort🎯 4 (Complex) | ⏱️ ~50 minutes The changes span multiple interconnected components with heterogeneous logic patterns: new FP8 kernel implementations with quantization-specific handling, per-expert weight stacking logic, runtime parameter extraction and propagation, and mlp_style branching logic. While individual files follow repetitive patterns (e.g., parameter handling), the kernel-level implementations and control-flow interactions require careful verification of numerical correctness, memory layout assumptions, and proper scale handling across quantized pathways. Pre-merge checks and finishing touches❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✨ Finishing touches
🧪 Generate unit tests (beta)
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. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 3
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (3)
tensorrt_llm/_torch/auto_deploy/transform/library/quantize_moe.py (1)
1-1: Add NVIDIA 2025 Apache-2.0 header.This Python source is missing the required header. Please prepend the standard 2025 NVIDIA SPDX header.
+ # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. + # SPDX-License-Identifier: Apache-2.0 + # + # Licensed under the Apache License, Version 2.0 (the "License"); + # you may not use this file except in compliance with the License. + # You may obtain a copy of the License at + # + # http://www.apache.org/licenses/LICENSE-2.0 + # + # Unless required by applicable law or agreed to in writing, software + # distributed under the License is distributed on an "AS IS" BASIS, + # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + # See the License for the specific language governing permissions and + # limitations under the License.tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py (1)
1-1: Add NVIDIA 2025 Apache-2.0 header.Please prepend the required header.
+ # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. + # SPDX-License-Identifier: Apache-2.0 + # + # Licensed under the Apache License, Version 2.0 (the "License"); + # you may not use this file except in compliance with the License. + # You may obtain a copy of the License at + # + # http://www.apache.org/licenses/LICENSE-2.0 + # + # Unless required by applicable law or agreed to in writing, software + # distributed under the License is distributed on an "AS IS" BASIS, + # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + # See the License for the specific language governing permissions and + # limitations under the License.tests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/triton_kernels/test_triton_moe.py (1)
1-3: Add NVIDIA 2025 Apache-2.0 header to test file.Tests are Python sources; add the header for compliance.
+ # SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved. + # SPDX-License-Identifier: Apache-2.0
🧹 Nitpick comments (5)
tests/integration/defs/accuracy/test_llm_api_autodeploy.py (2)
154-177: Align config shape with other suites (use transforms/compile_model).Match Llama3_1_8B/NemotronH patterns to avoid ignored keys.
- def get_default_kwargs(self): - return { - "skip_tokenizer_init": False, - "trust_remote_code": True, - # SSMs do not support cache reuse. - "kv_cache_config": { - "enable_block_reuse": False - }, - # Keep max_batch_size as in the PyTorch test to avoid OOM - "max_batch_size": 128, - # Model context length is 8K - "max_seq_len": 8192, - # Set explicitly to match default build_config behavior - "max_num_tokens": 8192, - "skip_loading_weights": False, - "compile_backend": "torch-cudagraph", - "free_mem_ratio": 0.7, - "cuda_graph_batch_sizes": [1, 2, 4, 8, 16, 32, 64, 128], - } + def get_default_kwargs(self): + return { + "skip_tokenizer_init": False, + "trust_remote_code": True, + "kv_cache_config": {"enable_block_reuse": False}, + "max_batch_size": 128, + "max_seq_len": 8192, + "max_num_tokens": 8192, + "skip_loading_weights": False, + "transforms": { + "resize_kv_cache": {"free_mem_ratio": 0.7}, + "compile_model": { + "backend": "torch-cudagraph", + "cuda_graph_batch_sizes": [1, 2, 4, 8, 16, 32, 64, 128], + }, + }, + }
186-197: Prefer decorator skip and drop unreachable body.Use @pytest.mark.skip to avoid executing dead code after pytest.skip().
- @pytest.mark.skip_less_device_memory(32000) - def test_auto_dtype(self): - pytest.skip("Nemotron-MOE is not in CI yet") - kwargs = self.get_default_kwargs() - sampling_params = self.get_default_sampling_params() - with AutoDeployLLM(model=self.MODEL_PATH, - tokenizer=self.MODEL_PATH, - **kwargs) as llm: - task = MMLU(self.MODEL_NAME) - task.evaluate(llm, sampling_params=sampling_params) - task = GSM8K(self.MODEL_NAME) - task.evaluate(llm) + @pytest.mark.skip_less_device_memory(32000) + @pytest.mark.skip(reason="Nemotron-MOE is not in CI yet") + def test_auto_dtype(self): + passOptionally validate that MODEL_PATH exists when running locally; the harness usually supports hub IDs. Do you want a helper that switches to hub if the local path is absent?
tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py (3)
51-55: Tiny simplification: avoid extra None-dim reshape.Direct indexing is clearer and avoids an unnecessary reshape.
- tokens_for_this_expert = x[None, top_x].reshape(-1, hidden_dim) + tokens_for_this_expert = x[top_x]
332-349: Silence unused args in fake kernel.Keep API stable but avoid ARG001 warnings by explicitly discarding.
def torch_quant_fp8_moe_fake( @@ - mlp_style: str = "gated_mlp", - act_fn: str = "silu", + mlp_style: str = "gated_mlp", + act_fn: str = "silu", ) -> torch.Tensor: - return torch.empty_like(x) + # Intentionally unused in fake; ensure lints stay quiet. + _ = (mlp_style, act_fn) + return torch.empty_like(x)
352-369: API parity suggestion: consider mlp_style/act_fn for NVFP4.For a consistent surface, accept mlp_style/act_fn (even if you only support "gated_mlp" for now) and validate early. This avoids branching in callers and mirrors FP8.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (7)
tensorrt_llm/_torch/auto_deploy/config/default.yaml(1 hunks)tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py(3 hunks)tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.py(5 hunks)tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py(2 hunks)tensorrt_llm/_torch/auto_deploy/transform/library/quantize_moe.py(1 hunks)tests/integration/defs/accuracy/test_llm_api_autodeploy.py(1 hunks)tests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/triton_kernels/test_triton_moe.py(1 hunks)
🧰 Additional context used
📓 Path-based instructions (3)
**/*.{h,hpp,hh,hxx,cpp,cxx,cc,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Use only spaces, no tabs; indent with 4 spaces.
Files:
tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.pytensorrt_llm/_torch/auto_deploy/transform/library/quantize_moe.pytensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.pytests/integration/defs/accuracy/test_llm_api_autodeploy.pytests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/triton_kernels/test_triton_moe.pytensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py
**/*.py
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
**/*.py: Python code must target Python 3.8+.
Indent Python code with 4 spaces; do not use tabs.
Maintain module namespace when importing; prefer 'from package.subpackage import foo' then 'foo.SomeClass()' instead of importing the class directly.
Python filenames should be snake_case (e.g., some_file.py).
Python classes use PascalCase names.
Functions and methods use snake_case names.
Local variables use snake_case; prefix 'k' for variables that start with a number (e.g., k_99th_percentile).
Global variables use upper SNAKE_CASE prefixed with 'G' (e.g., G_MY_GLOBAL).
Constants use upper SNAKE_CASE (e.g., MY_CONSTANT).
Avoid shadowing variables from an outer scope.
Initialize all externally visible members of a class in the constructor.
Prefer docstrings for interfaces that may be used outside a file; comments for in-function or file-local interfaces.
Use Google-style docstrings for classes and functions (Sphinx-parsable).
Document attributes and variables inline so they render under the class/function docstring.
Avoid reflection when a simpler, explicit approach suffices (e.g., avoid dict(**locals()) patterns).
In try/except, catch the most specific exceptions possible.
For duck-typing try/except, keep the try body minimal and use else for the main logic.
Files:
tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.pytensorrt_llm/_torch/auto_deploy/transform/library/quantize_moe.pytensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.pytests/integration/defs/accuracy/test_llm_api_autodeploy.pytests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/triton_kernels/test_triton_moe.pytensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py
**/*.{cpp,cxx,cc,h,hpp,hh,hxx,cu,cuh,py}
📄 CodeRabbit inference engine (CODING_GUIDELINES.md)
Prepend the NVIDIA Apache-2.0 copyright header with current year to the top of all source files (e.g., .cpp, .h, .cu, .py).
Files:
tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.pytensorrt_llm/_torch/auto_deploy/transform/library/quantize_moe.pytensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.pytests/integration/defs/accuracy/test_llm_api_autodeploy.pytests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/triton_kernels/test_triton_moe.pytensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py
🧬 Code graph analysis (4)
tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py (5)
tensorrt_llm/_torch/auto_deploy/utils/node_utils.py (2)
is_op(179-202)extract_op_args(407-444)tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py (1)
torch_quant_fp8_moe(218-329)tensorrt_llm/module.py (1)
register_parameter(186-190)tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.py (2)
triton_quant_fp8_moe(521-607)triton_quant_fp8_moe(611-627)tensorrt_llm/_torch/auto_deploy/transform/interface.py (4)
TransformRegistry(503-531)register(509-516)BaseTransform(213-500)TransformInfo(121-174)
tensorrt_llm/_torch/auto_deploy/transform/library/quantize_moe.py (1)
tensorrt_llm/_torch/auto_deploy/utils/pattern_matcher.py (1)
call_function(249-276)
tests/integration/defs/accuracy/test_llm_api_autodeploy.py (2)
tests/integration/defs/accuracy/accuracy_core.py (5)
LlmapiAccuracyTestHarness(846-857)MMLU(317-331)evaluate(184-247)evaluate(765-775)GSM8K(334-349)tensorrt_llm/_torch/auto_deploy/models/factory.py (2)
model(125-127)tokenizer(130-132)
tests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/triton_kernels/test_triton_moe.py (2)
tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.py (2)
triton_quant_fp8_moe(521-607)triton_quant_fp8_moe(611-627)tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py (1)
torch_quant_fp8_moe(218-329)
🪛 Ruff (0.14.1)
tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py
612-613: try-except-continue detected, consider logging the exception
(S112)
612-612: Do not catch blind exception: Exception
(BLE001)
772-772: Unused method argument: cm
(ARG002)
773-773: Unused method argument: factory
(ARG002)
774-774: Unused method argument: shared_config
(ARG002)
tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.py
354-354: Unused function argument: num_tokens_post_padded
(ARG001)
423-423: Avoid specifying long messages outside the exception class
(TRY003)
505-505: Unused function argument: selected_experts
(ARG001)
506-506: Unused function argument: routing_weights
(ARG001)
507-507: Unused function argument: w1_stacked_weight
(ARG001)
508-508: Unused function argument: w2_stacked_weight
(ARG001)
527-527: Unused function argument: w3_weight
(ARG001)
530-530: Unused function argument: w3_input_scale
(ARG001)
533-533: Unused function argument: w3_weight_scale
(ARG001)
535-535: Unused function argument: act_fn
(ARG001)
613-613: Unused function argument: selected_experts
(ARG001)
614-614: Unused function argument: routing_weights
(ARG001)
615-615: Unused function argument: w1_weight
(ARG001)
616-616: Unused function argument: w2_weight
(ARG001)
617-617: Unused function argument: w3_weight
(ARG001)
618-618: Unused function argument: w1_input_scale
(ARG001)
619-619: Unused function argument: w2_input_scale
(ARG001)
620-620: Unused function argument: w3_input_scale
(ARG001)
621-621: Unused function argument: w1_weight_scale
(ARG001)
622-622: Unused function argument: w2_weight_scale
(ARG001)
623-623: Unused function argument: w3_weight_scale
(ARG001)
624-624: Unused function argument: mlp_style
(ARG001)
625-625: Unused function argument: act_fn
(ARG001)
tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/torch_moe.py
327-327: Avoid specifying long messages outside the exception class
(TRY003)
346-346: Unused function argument: mlp_style
(ARG001)
347-347: Unused function argument: act_fn
(ARG001)
⏰ 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). (1)
- GitHub Check: Pre-commit Check
🔇 Additional comments (15)
tests/unittest/_torch/auto_deploy/unit/singlegpu/custom_ops/triton_kernels/test_triton_moe.py (1)
220-360: Test function verified as single definition; parity test setup is solid.Verification confirms only one definition of
test_triton_quant_fp8_moe_matches_torch_quant_fp8_moeexists in the file (line 220). The FP8 quantization parity test is well-structured with proper controlled routing, appropriate per-expert weight scaling, and reasonable tolerances for FP8 comparisons.tensorrt_llm/_torch/auto_deploy/config/default.yaml (1)
119-121: Verification confirms transform is safely registered with proper no-op behavior.The
fuse_fp8_moetransform is registered attensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py:762and implements graceful skipping: when_stack_fp8_moe_weights()finds no FP8 MoE weights (counter == 0), it returnsTransformInfo(skipped=True). Safe to enable by default.tensorrt_llm/_torch/auto_deploy/transform/library/fused_moe.py (6)
616-627: LGTM - Good helper for retrieving parameters or buffers.The helper correctly handles both direct attributes and nested submodule attributes, which is necessary since scales may be registered as buffers rather than parameters.
630-665: LGTM - Correct handling of optional w3 weights and scales.The stacking logic correctly handles both gated MLP (with w3) and standard MLP (without w3) styles by creating empty tensors when w3_list is empty. This maintains consistent tensor shapes across different MLP styles.
681-704: Verify that scales should be registered as parameters rather than buffers.The original scales were likely registered as buffers (non-trainable tensors), but here they're being registered as parameters with
requires_grad=False. While functionally similar for inference, this changes their semantic meaning and how they're handled by PyTorch's module system.Consider whether these should be registered as buffers instead:
gm.register_buffer(new_key_w1_input_scale, w1_input_scale_stacked)If the distinction matters for serialization, state dict handling, or other framework features, please verify the correct approach. If parameters are intentional, consider adding a comment explaining why scales are parameters rather than buffers.
707-728: LGTM - Correct graph node replacement.The node replacement correctly:
- Uses
graph.get_attr()to reference the newly registered stacked parameters- Preserves the original node's
kwargs(e.g.,mlp_style,act_fn)- Replaces all uses before erasing the old node
730-737: LGTM - Proper cleanup of unused parameters.The cleanup correctly removes dead code and unused submodules after stacking, which is essential for memory efficiency when dealing with large models.
762-785: LGTM - Clean transform implementation following established patterns.The
FuseFP8Moetransform correctly:
- Follows the same pattern as
FuseMoefor consistency- Uses
cuda_memory_trackerto monitor memory during the transformation- Sets
skipped=Truewhen no FP8 MoE patterns are found- Implements the
BaseTransforminterface (unused arguments are expected)tensorrt_llm/_torch/auto_deploy/custom_ops/fused_moe/triton_moe.py (7)
347-412: LGTM - Clean refactoring with proper kernel dispatch.The refactoring correctly:
- Consolidates common arguments between unquantized and FP8 paths
- Uses conditional dispatch based on scale presence to select the appropriate kernel
- Maintains backward compatibility for the unquantized path
Note: The static analysis warning about
num_tokens_post_paddedbeing unused is a false positive—it's used incommon_argsat line 384.
414-423: LGTM - Correct dtype mapping.The helper correctly maps PyTorch dtypes to Triton compute types. The function is simple and handles the common FP8 use cases (BF16/FP16 output).
426-481: LGTM - Well-factored internal MoE implementation.The refactored
_fused_moe_mlp_relu2function correctly:
- Handles token packing and kernel configuration
- Performs two GEMMs with ReLU² activation in between
- Uses the new
_invoke_kernelinterface for both passes
513-517: LGTM - Correct FP8 quantization with clamping.The quantization helper correctly clamps values to the FP8 E4M3 range before conversion, which is necessary to prevent overflow and matches the behavior of
torch_quant_fp8_linear.
534-539: Document current limitations of FP8 MoE implementation.The FP8 Triton MoE currently only supports:
mlp_style=="mlp"(2-layer non-gated MLP)- ReLU² activation (implicit, hardcoded)
However, the function signature accepts
mlp_styleandact_fnparameters to match the interface oftorch_quant_fp8_moe. Theact_fnparameter is unused.Consider:
- Adding a check for
act_fnto ensure it matches the hardcoded ReLU² behavior- Updating the docstring to clearly document these limitations
- Adding a TODO comment for future gated MLP support
"""Triton FP8 W8A8 MoE with 2-layer MLP and ReLU^2 activation.""" if mlp_style != "mlp": raise NotImplementedError("triton_quant_fp8_moe currently supports mlp_style=='mlp' only") + + # Currently only ReLU^2 activation is supported (hardcoded in kernel) + if act_fn not in ("relu2", "silu"): # Accept silu for compatibility but treat as relu2 + raise NotImplementedError(f"triton_quant_fp8_moe currently only supports relu2 activation, got {act_fn}")
549-550: Verify assumption that all experts share the same input scale.Lines 549-550 extract only the first element of
w1_input_scaleandw2_input_scale, assuming all experts use the same input scale:a1_scale = w1_input_scale[0].to(torch.float32).reshape(1).contiguous() a2_scale = w2_input_scale[0].to(torch.float32).reshape(1).contiguous()This differs from the weight scales (b1_scale, b2_scale) which are per-expert. Is this intentional?
If input quantization scales are truly shared across all experts (i.e., the input tensor has a single scale regardless of which expert processes it), this is correct. However, please verify this matches the quantization strategy used in the model and is consistent with how
torch_quant_fp8_moein torch_moe.py handles per-expert input scales.If input scales should be per-expert, the kernel would need to be updated to load the appropriate scale based on
off_experts.
610-627: LGTM - Fake implementation follows PyTorch custom op pattern.The fake implementation correctly returns an output tensor with the same shape as the input, which is required for PyTorch's tracing and shape inference. The unused argument warnings from static analysis are expected for fake implementations.
Signed-off-by: Chenghao Zhang <[email protected]>
|
/bot run |
|
PR_Github #22245 [ run ] triggered by Bot. Commit: |
|
PR_Github #22245 [ run ] completed with state |
|
/bot run |
|
PR_Github #22266 [ run ] triggered by Bot. Commit: |
|
PR_Github #22266 [ run ] completed with state |
Signed-off-by: Fridah-nv <[email protected]>
|
/bot run |
|
PR_Github #22330 [ run ] triggered by Bot. Commit: |
|
PR_Github #22330 [ run ] completed with state |
Signed-off-by: nvchenghaoz <[email protected]>
|
/bot run |
|
PR_Github #22369 [ run ] triggered by Bot. Commit: |
|
PR_Github #22369 [ run ] completed with state |
|
/bot run |
|
PR_Github #22460 [ run ] triggered by Bot. Commit: |
Fridah-nv
left a comment
Fridah-nv
left a comment
There was a problem hiding this comment.
Rubber stamped per request.
|
/bot run |
|
/bot kill |
|
PR_Github #22478 [ run ] triggered by Bot. Commit: |
|
PR_Github #22460 [ run ] completed with state |
|
/bot run |
|
PR_Github #22479 [ kill ] triggered by Bot. Commit: |
|
PR_Github #22478 [ run ] completed with state |
|
PR_Github #22479 [ kill ] completed with state |
|
PR_Github #22480 [ run ] triggered by Bot. Commit: |
|
PR_Github #22480 [ run ] completed with state |
Signed-off-by: nvchenghaoz <[email protected]>
|
/bot run |
|
PR_Github #22489 [ run ] triggered by Bot. Commit: |
|
PR_Github #22489 [ run ] completed with state |
|
/bot run |
|
PR_Github #22508 [ run ] triggered by Bot. Commit: |
|
PR_Github #22508 [ run ] completed with state |
Signed-off-by: Chenghao Zhang <[email protected]> Signed-off-by: Fridah-nv <[email protected]> Signed-off-by: nvchenghaoz <[email protected]> Co-authored-by: Suyog Gupta <[email protected]> Co-authored-by: Fridah-nv <[email protected]>
Signed-off-by: Chenghao Zhang <[email protected]> Signed-off-by: Fridah-nv <[email protected]> Signed-off-by: nvchenghaoz <[email protected]> Co-authored-by: Suyog Gupta <[email protected]> Co-authored-by: Fridah-nv <[email protected]>
Signed-off-by: Chenghao Zhang <[email protected]> Signed-off-by: Fridah-nv <[email protected]> Signed-off-by: nvchenghaoz <[email protected]> Co-authored-by: Suyog Gupta <[email protected]> Co-authored-by: Fridah-nv <[email protected]>
Signed-off-by: Chenghao Zhang <[email protected]> Signed-off-by: Fridah-nv <[email protected]> Signed-off-by: nvchenghaoz <[email protected]> Co-authored-by: Suyog Gupta <[email protected]> Co-authored-by: Fridah-nv <[email protected]>
4 participants
Summary by CodeRabbit
Release Notes
New Features
Tests