[None] [feat] Optimize the algorithm part of RocketKV #9333
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Contributor
📝 WalkthroughWalkthroughThe PR updates RocketKV sparse attention by replacing the top-r parameter with top-k, introducing KT cache data type selection, and refactoring kernels to use block-oriented indexing. Configuration defaults are adjusted across multiple files, and new parameters Changes
Sequence Diagram(s)sequenceDiagram
participant CLI as CLI Arguments
participant Config as RocketSparseAttentionConfig
participant KVCache as KvCacheConfig
participant KVMgr as RocketKVCacheManager
participant Kernel as Sparse Kernels
Note over CLI: --topk, --kt_cache_dtype, --tokens_per_block
CLI->>Config: topk (int, default 64)
CLI->>Config: kt_cache_dtype (str, default 'float8_e5m2')
CLI->>KVCache: tokens_per_block (int, default 64)
Config->>KVMgr: Pass topk, kt_cache_dtype
KVCache->>KVMgr: Pass tokens_per_block
KVMgr->>KVMgr: Set KT cache buffer dtype based on kt_cache_dtype
KVMgr->>Kernel: total_rocket_k_ctx_tokens, prompt_budget
Note over Kernel: Block-oriented indexing (replaces dim_pos)
Kernel->>Kernel: Grid config: (batch_size, num_kv_heads, num_KT_token_tiles)
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes
Pre-merge checks and finishing touches❌ Failed checks (2 warnings)
✅ Passed checks (1 passed)
✨ Finishing touches
🧪 Generate unit tests (beta)
Tip 📝 Customizable high-level summaries are now available in beta!You can now customize how CodeRabbit generates the high-level summary in your pull requests — including its content, structure, tone, and formatting.
Example instruction:
Note: This feature is currently in beta for Pro-tier users, and pricing will be announced later. 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}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 4
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (2)
tensorrt_llm/llmapi/llm_args.py (1)
1440-1516: Add validation forKvCacheConfig.tokens_per_blockto prevent invalid KV cache layouts
tokens_per_blockis the logical capacity of a KV cache block and is used in block-count computations (and in downstream logic like paged KT cache / attention windows). If a caller sets it to0or a negative value, you will eventually hit division-by-zero or out‑of‑range indexing in those paths.It’s worth enforcing a positive value at the config level, similar to the existing validators on
free_gpu_memory_fractionandmax_gpu_total_bytes. For example:class KvCacheConfig(StrictBaseModel, PybindMirror): @@ - tokens_per_block: int = Field(default=32, - description="The number of tokens per block.") + tokens_per_block: int = Field( + default=32, + description="The number of tokens per block.") @@ @field_validator('max_attention_window') @classmethod def validate_max_attention_window(cls, v: Optional[List[int]]): @@ return v + + @field_validator('tokens_per_block') + @classmethod + def validate_tokens_per_block(cls, v: int) -> int: + if v <= 0: + raise ValueError( + "kv_cache_config.tokens_per_block must be a positive integer") + return vThis keeps misconfigurations from propagating into the KV cache manager where they’re harder to diagnose. Based on learnings
tensorrt_llm/_torch/attention_backend/sparse/kernel.py (1)
894-981: Critical: Mismatch in KT token calculation between wrapper and kernel.The wrapper function calculates grid dimension z using
prompt_budgetbut the kernel recalculatestotal_kt_tokensper-batch usingnum_sparse_tokensfrom sparse offsets. These can differ, causing the grid to be sized incorrectly relative to actual token counts:
- Wrapper (kernel.py:1018-1020):
total_kt_tokens = (prompt_budget + kt_page_size - 1) // kt_page_size- Kernel (kernel.py:923):
total_kt_tokens = (num_sparse_tokens + kt_page_size - 1) // kt_page_sizeWhen per-batch
num_sparse_tokens < prompt_budget, the kernel mask at line 925 uses the smaller value while grid.z was sized for the larger value. This creates a correctness risk.Additionally, no test coverage exists for this kernel. The search found tests for other sparse kernels (test_triton_bmm.py, test_triton_topk.py, test_flash_mla.py) but nothing for
rocket_update_kt_cache_ctx.Fix the calculation to use a consistent total_kt_tokens value, and add comprehensive test coverage for edge cases: partial pages, varying batch sizes, and different head dimensions.
🧹 Nitpick comments (7)
tensorrt_llm/llmapi/llm_args.py (1)
222-240: RocketSparseAttentionConfig RocketKV knobs look consistent; small API polish possibleThe new defaults (
window_size=32,kernel_size=63,topr=128,topk=64,prompt_budget=2048) and the addedpage_size/kt_cache_dtypefields are coherent with the example scripts and tests, andsupports_backendcorrectly limits this to the PyTorch backend.Two optional improvements:
- The
page_sizename is very close toKvCacheConfig.tokens_per_block. A brief clarification in thepage_sizedescription that it is the KT index page size (not KV cache block size) would help avoid confusion now that a separatetokens_per_blockknob exists.- For
kt_cache_dtype, consider tightening the type to aLiteral['bfloat16', 'float8_e5m2']or an Enum instead ofOptional[str]pluschoices, so that Pydantic and static tooling enforce valid values at type level.tests/unittest/_torch/attention/sparse/test_triton_bmm.py (1)
183-216: Reference paged KT BMM matches new layout; minor cleanups are optionalThe reference now correctly produces
scoreswith shape[num_kv_heads, num_heads_per_kv, total_kt_tokens]and usesoutput_offset = batch_idx * max_kt_tokens, which aligns with howoutput_offsets/maskare computed in the test, so the layout/indexing look sound.A couple of non‑essential cleanups you could consider:
k_vecandk_selecteddepend only onkv_head_idx, but are recomputed inside theq_head_idxloop; they could be moved one level up.- Given you intentionally use the same tensor for min/max,
k_selected = torch.where(dim_pos_vec, k_vec, k_vec)is a no‑op; keeping thedim_pos_veccomment for documentation but dropping thewherecall would reduce confusion.These are purely cosmetic in a test‑only reference path.
examples/llm-api/llm_sparse_attention.py (1)
68-76: RocketKV CLI knobs are wired correctly; consider clarifying KT vs KV dtypes and tokens_per_blockThe new arguments
--topk,--kt_cache_dtype, and--tokens_per_blockare correctly plumbed:
topk/kt_cache_dtypego intoRocketSparseAttentionConfig, matching its updated signature.tokens_per_blockandkv_cache_dtypego intoKvCacheConfig, matching the Pydantic model.Two small clarity improvements you might consider:
- In the help text for
--kt_cache_dtype, call out explicitly that it controls the KT cache precision and is independent of--kv_cache_dtype, which still sets the KV cache dtype.- For
--tokens_per_block, a short note that this affects the PyTorch backend’s KV cache manager page/block size (and is ignored by the TensorRT backend) would help users understand when the knob has an effect.Behavior-wise this looks good as is.
Also applies to: 117-117, 143-149, 193-200
examples/longbench/eval_longbench_v1.py (1)
153-162: LongBench v1 RocketKV knobs are correctly integrated; only minor doc consistency nitsHere,
--topk,--kt_cache_dtype, and--tokens_per_blockare all correctly threaded:
topk/kt_cache_dtypeare passed intoRocketSparseAttentionConfigwhen--rocket_sparseis enabled.tokens_per_blockis passed intoKvCacheConfig, aligning with the new KV cache block‑size knob.If you want to polish further (optional):
- Mirror the
--kt_cache_dtypehelp string fromllm_sparse_attention.pyto keep CLI UX consistent across examples.- Consider mentioning in the
--tokens_per_blockhelp that this primarily affects the PyTorch backend’s KV cache manager so users don’t expect it to change prebuilt TensorRT engines.Functionally this looks solid.
Also applies to: 168-173, 325-330, 341-347
tensorrt_llm/_torch/attention_backend/sparse/rocket.py (2)
462-465: Verify contiguity check performance impact.The contiguity checks with conditional
.contiguous()calls are defensive but create new tensors if the input is non-contiguous. If this method is called frequently in the hot path, consider whether upstream callers should be required to provide contiguous tensors to avoid allocation overhead.Based on learnings: TensorRT-LLM's attention backend typically trusts callers to provide correctly formatted tensors without validation. Consider documenting the contiguity requirement instead of runtime checks, or profile to confirm the overhead is negligible.
625-625: Restrictive assertion may limit future extensibility.The hard assertion
assert sparse_attention_config.kt_cache_dtype == 'bfloat16'prevents Vanilla RocketKV from supporting FP8 KT cache. If FP8 support for Vanilla is planned or could be useful, consider replacing this with a capability check or warning.- assert sparse_attention_config.kt_cache_dtype == 'bfloat16', "Only bfloat16 kt cache is supported for Vanilla RocketKV" + if sparse_attention_config.kt_cache_dtype != 'bfloat16': + raise NotImplementedError(f"Vanilla RocketKV currently only supports bfloat16 kt_cache, got {sparse_attention_config.kt_cache_dtype}")This makes the temporary limitation clearer and more maintainable.
tensorrt_llm/_torch/attention_backend/sparse/kernel.py (1)
1195-1195: Static analysis: Replace lambda with def.Ruff flags the lambda expression as a style issue. While this works, following the style guideline improves readability.
Apply this diff:
- grid = lambda meta: (num_gen_tokens, num_kv_heads) + def grid(meta): + return (num_gen_tokens, num_kv_heads)However, since
metais unused, you could also use:- grid = lambda meta: (num_gen_tokens, num_kv_heads) + grid = (num_gen_tokens, num_kv_heads)if the kernel invocation accepts a tuple directly.
📜 Review details
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (8)
examples/llm-api/llm_sparse_attention.py(4 hunks)examples/longbench/eval_longbench_v1.py(3 hunks)examples/longbench/eval_longbench_v2.py(2 hunks)tensorrt_llm/_torch/attention_backend/sparse/kernel.py(13 hunks)tensorrt_llm/_torch/attention_backend/sparse/rocket.py(10 hunks)tensorrt_llm/llmapi/llm_args.py(1 hunks)tests/unittest/_torch/attention/sparse/test_rocketkv.py(5 hunks)tests/unittest/_torch/attention/sparse/test_triton_bmm.py(2 hunks)
🧰 Additional context used
🧠 Learnings (10)
📓 Common learnings
Learnt from: thorjohnsen
Repo: NVIDIA/TensorRT-LLM PR: 6910
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-14T21:04:50.248Z
Learning: In KV cache onboarding logic during prefill in cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, when calculating which blocks fall within the attention window, use getTokensPerBlock() to advance token indices rather than block->getUniqueTokens().size(), because the calculation needs to consider the post-prefill state where blocks will be filled to capacity, not their current token count.
Learnt from: eopXD
Repo: NVIDIA/TensorRT-LLM PR: 6767
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-15T06:46:54.897Z
Learning: In cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp addToken function, newly allocated blocks are unshared by design. The beam search path in addToken (when sequence.getNumTokens() > windowSize) is currently broken/non-functional with SWA, so the block allocation doesn't follow a shared-then-unshared pattern.
Learnt from: eopXD
Repo: NVIDIA/TensorRT-LLM PR: 6767
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-15T06:46:53.813Z
Learning: In the TensorRT-LLM KV cache manager, SWA (Sliding Window Attention) combined with beam search is currently in a broken/non-functional state and is planned for future rework. During preparatory refactoring phases, code related to SWA+beam search may intentionally remain in a non-working state until the broader rework is completed.
📚 Learning: 2025-08-14T15:38:01.771Z
Learnt from: MatthiasKohl
Repo: NVIDIA/TensorRT-LLM PR: 6904
File: cpp/tensorrt_llm/pybind/thop/bindings.cpp:55-57
Timestamp: 2025-08-14T15:38:01.771Z
Learning: In TensorRT-LLM Python bindings, tensor parameter collections like mla_tensor_params and spec_decoding_tensor_params are kept as required parameters without defaults to maintain API consistency, even when it might affect backward compatibility.
Applied to files:
tensorrt_llm/llmapi/llm_args.py
📚 Learning: 2025-08-14T21:04:50.248Z
Learnt from: thorjohnsen
Repo: NVIDIA/TensorRT-LLM PR: 6910
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-14T21:04:50.248Z
Learning: In KV cache onboarding logic during prefill in cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, when calculating which blocks fall within the attention window, use getTokensPerBlock() to advance token indices rather than block->getUniqueTokens().size(), because the calculation needs to consider the post-prefill state where blocks will be filled to capacity, not their current token count.
Applied to files:
tests/unittest/_torch/attention/sparse/test_rocketkv.pyexamples/longbench/eval_longbench_v1.pyexamples/llm-api/llm_sparse_attention.pytests/unittest/_torch/attention/sparse/test_triton_bmm.pytensorrt_llm/_torch/attention_backend/sparse/rocket.pyexamples/longbench/eval_longbench_v2.pytensorrt_llm/_torch/attention_backend/sparse/kernel.py
📚 Learning: 2025-08-26T09:37:10.463Z
Learnt from: jiaganc
Repo: NVIDIA/TensorRT-LLM PR: 7031
File: tensorrt_llm/bench/dataclasses/configuration.py:90-104
Timestamp: 2025-08-26T09:37:10.463Z
Learning: In TensorRT-LLM's bench configuration, the `get_pytorch_perf_config()` method returns `self.pytorch_config` which is a Dict[str, Any] that can contain default values including `cuda_graph_config`, making the fallback `llm_args["cuda_graph_config"]` safe to use.
Applied to files:
tests/unittest/_torch/attention/sparse/test_rocketkv.py
📚 Learning: 2025-08-14T15:43:23.107Z
Learnt from: MatthiasKohl
Repo: NVIDIA/TensorRT-LLM PR: 6904
File: tensorrt_llm/_torch/attention_backend/trtllm.py:259-262
Timestamp: 2025-08-14T15:43:23.107Z
Learning: In TensorRT-LLM's attention backend, tensor parameters in the plan() method are assigned directly without validation (dtype, device, contiguity checks). This maintains consistency across all tensor inputs and follows the pattern of trusting callers to provide correctly formatted tensors.
Applied to files:
tests/unittest/_torch/attention/sparse/test_rocketkv.pytensorrt_llm/_torch/attention_backend/sparse/kernel.py
📚 Learning: 2025-08-15T06:46:54.897Z
Learnt from: eopXD
Repo: NVIDIA/TensorRT-LLM PR: 6767
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-15T06:46:54.897Z
Learning: In cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp addToken function, newly allocated blocks are unshared by design. The beam search path in addToken (when sequence.getNumTokens() > windowSize) is currently broken/non-functional with SWA, so the block allocation doesn't follow a shared-then-unshared pattern.
Applied to files:
examples/llm-api/llm_sparse_attention.pytensorrt_llm/_torch/attention_backend/sparse/kernel.py
📚 Learning: 2025-09-23T14:58:05.372Z
Learnt from: nv-lschneider
Repo: NVIDIA/TensorRT-LLM PR: 7910
File: cpp/tensorrt_llm/kernels/nccl_device/config.cu:42-49
Timestamp: 2025-09-23T14:58:05.372Z
Learning: In TensorRT-LLM NCCL device kernels (cpp/tensorrt_llm/kernels/nccl_device/), the token partitioning intentionally uses ceil-like distribution (same token_per_rank for all ranks) to ensure all ranks launch the same number of blocks. This is required for optimal NCCL device API barrier performance, even though it may launch extra blocks for non-existent tokens on later ranks. Runtime bounds checking in the kernel (blockID validation) handles the overshoot cases.
Applied to files:
tensorrt_llm/_torch/attention_backend/sparse/kernel.py
📚 Learning: 2025-08-15T06:46:53.813Z
Learnt from: eopXD
Repo: NVIDIA/TensorRT-LLM PR: 6767
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:0-0
Timestamp: 2025-08-15T06:46:53.813Z
Learning: In the TensorRT-LLM KV cache manager, SWA (Sliding Window Attention) combined with beam search is currently in a broken/non-functional state and is planned for future rework. During preparatory refactoring phases, code related to SWA+beam search may intentionally remain in a non-working state until the broader rework is completed.
Applied to files:
tensorrt_llm/_torch/attention_backend/sparse/kernel.py
📚 Learning: 2025-08-22T01:54:35.850Z
Learnt from: djns99
Repo: NVIDIA/TensorRT-LLM PR: 7104
File: cpp/tensorrt_llm/kernels/cutlass_kernels/include/moe_kernels.h:999-1000
Timestamp: 2025-08-22T01:54:35.850Z
Learning: The `internal_cutlass_kernels` directory in TensorRT-LLM is a mirror of an internal NVIDIA repository and maintains its own implementation and API that may diverge from the public `cutlass_kernels` version. API inconsistencies between these two directories are intentional and by design, not bugs to be fixed.
Applied to files:
tensorrt_llm/_torch/attention_backend/sparse/kernel.py
📚 Learning: 2025-08-21T09:41:49.347Z
Learnt from: eopXD
Repo: NVIDIA/TensorRT-LLM PR: 6768
File: cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp:2010-2045
Timestamp: 2025-08-21T09:41:49.347Z
Learning: In cpp/tensorrt_llm/batch_manager/kvCacheManager.cpp, updateSequenceCacheBlockOffsets is specifically for updating bookkeeping when blocks are added during the context phase, not for refreshing offsets after detach operations. During detach operations, GenerationRequest::removeFrontBlock handles the necessary cache block bookkeeping internally.
Applied to files:
tensorrt_llm/_torch/attention_backend/sparse/kernel.py
🧬 Code graph analysis (6)
tensorrt_llm/llmapi/llm_args.py (1)
tensorrt_llm/_torch/attention_backend/flashinfer.py (1)
page_size(200-204)
tests/unittest/_torch/attention/sparse/test_rocketkv.py (1)
tensorrt_llm/llmapi/llm_args.py (1)
RocketSparseAttentionConfig(222-250)
examples/longbench/eval_longbench_v1.py (2)
tensorrt_llm/_torch/attention_backend/trtllm.py (1)
tokens_per_block(654-658)tensorrt_llm/functional.py (1)
topk(7318-7414)
examples/llm-api/llm_sparse_attention.py (2)
tensorrt_llm/_torch/attention_backend/trtllm.py (1)
tokens_per_block(654-658)tensorrt_llm/functional.py (1)
topk(7318-7414)
tensorrt_llm/_torch/attention_backend/sparse/rocket.py (3)
tensorrt_llm/_torch/attention_backend/flashinfer.py (1)
page_size(200-204)tensorrt_llm/_torch/attention_backend/trtllm.py (1)
tokens_per_block(654-658)tensorrt_llm/_torch/attention_backend/sparse/dsa.py (1)
get_cache_bytes_per_token(1569-1589)
tensorrt_llm/_torch/attention_backend/sparse/kernel.py (2)
tensorrt_llm/_torch/attention_backend/trtllm.py (1)
tokens_per_block(654-658)cpp/tensorrt_llm/kernels/sparseAttentionKernels.h (1)
sparse_kv_offsets(33-85)
🪛 Ruff (0.14.5)
tensorrt_llm/_torch/attention_backend/sparse/kernel.py
1195-1195: Do not assign a lambda expression, use a def
Rewrite grid as a def
(E731)
1195-1195: Unused lambda argument: meta
(ARG005)
⏰ 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 (7)
tests/unittest/_torch/attention/sparse/test_rocketkv.py (2)
374-386: Verify intentional differences in test configurations.The test uses different configurations for vanilla vs trtllm:
page_size=2for bothtopk=128for vanilla,topk=64for trtllmtopr=96for bothThese differences may be intentional for testing different code paths, but please confirm this is the intended test design rather than an oversight.
468-473: FP8 initialization loses statistical properties without explicit scaling.The naive conversion from float16 to float8_e5m2 does not preserve the normal distribution's properties. Quantization noise is large due to FP8 having only 2 mantissa bits and can change the sample mean and especially variance/tails. Per-tensor or per-channel scaling should be used when converting to FP8 to reduce bias and clipping.
The current implementation (lines 468-473) should either:
- Apply explicit per-tensor scaling before the conversion, or
- Use an alternative initialization approach that directly supports float8_e5m2, or
- Document why unscaled conversion is acceptable for this specific test
⛔ Skipped due to learnings
Learnt from: jhaotingc Repo: NVIDIA/TensorRT-LLM PR: 7856 File: cpp/tensorrt_llm/thop/fp8BlockScaleMoe.cpp:159-166 Timestamp: 2025-09-19T21:28:13.751Z Learning: In TensorRT-LLM blockScaleMoe routing (cpp/tensorrt_llm/kernels/trtllmGenKernels/blockScaleMoe/runner.cu), the DeepSeek routing method performs reinterpret_cast<float*>(routingLogits) at line 89, which could cause issues if routing_logits are BF16. However, Qwen3-FP8 models use RenormalizeNaive routing method and are not affected by this dtype casting issue.Learnt from: nvchenghaoz Repo: NVIDIA/TensorRT-LLM PR: 8469 File: tensorrt_llm/_torch/auto_deploy/transform/library/rms_norm.py:180-182 Timestamp: 2025-10-20T17:09:21.560Z Learning: In tensorrt_llm/_torch/auto_deploy/transform/library/rms_norm.py, the _gated_rmsnorm_replacement function does not need to cast the output of torch.ops.auto_deploy.torch_rmsnorm_gated back to the input dtype, even though the custom op returns fp32. The dtype handling is managed elsewhere or the fp32 output is acceptable for downstream consumers.tensorrt_llm/_torch/attention_backend/sparse/rocket.py (2)
48-49: Good validation: page_size must be power of 2.This assertion enforces a critical constraint that the block-oriented indexing relies on. The error will catch misconfigurations early.
258-267: Properly handles empty valid batch case.Setting
max_rocket_k_ctx_len=0andtotal_rocket_k_ctx_tokens=0whenvalid_batch_size=0prevents undefined behavior from calling.max().item()on an empty tensor slice.tensorrt_llm/_torch/attention_backend/sparse/kernel.py (3)
508-520: Softmax kernel generalized for 2D and 3D inputs.The updated logic correctly handles both input shapes:
- 2D:
[num_heads, total_k_tokens]withlen_per_seq=1- 3D:
[num_heads, len_per_seq, total_k_tokens]The grid calculation
(num_heads, batch_size * len_per_seq)properly accounts for both cases.
1062-1152: Review comment is unsupported by evidence.The
dim_poscomputation viatl.sum(q_values, axis=0) > 0is mathematically equivalent to and directly validated by the test reference (test_triton_bmm.py lines 200-201), which implements the identical logic. The test confirms this is the expected behavior, not a deviation.The claim of a "critical logic change from the previous dim_pos-based approach" lacks supporting evidence. The current implementation IS the dim_pos-based approach and aligns perfectly with test expectations. No alternative previous implementation was found in the codebase.
The shape analysis (Q: [Q_BLOCK_SIZE, DIM_BLOCK_SIZE], KT: [DIM_BLOCK_SIZE, KT_BLOCK_SIZE], output: [Q_BLOCK_SIZE, KT_BLOCK_SIZE]) and output indexing logic are correct as stated.
Likely an incorrect or invalid review comment.
1190-1197: Output shape mismatch breaks downstream function compatibility.The function returns
[num_kv_heads, num_heads_per_kv, total_kt_tokens], buttriton_rocket_reduce_scores(rocket.py:528) expects[num_kv_heads * num_heads_per_kv, 1, total_kt_tokens]. The dimensions are incompatible and no reshape operation exists between the calls (rocket.py:512-528). A reshape is needed before passing scores totriton_rocket_reduce_scores, or the function should flatten the first two dimensions before returning.⛔ Skipped due to learnings
Learnt from: sklevtsov-nvidia Repo: NVIDIA/TensorRT-LLM PR: 3294 File: cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_tma_warp_specialized_input.cu:118-127 Timestamp: 2025-08-09T20:57:04.084Z Learning: In the CUTLASS MoE finalize fusion implementation (cpp/tensorrt_llm/kernels/cutlass_kernels/moe_gemm/moe_gemm_tma_warp_specialized_input.cu), when setting `fused_finalize_epilogue.stride_final_output` with shape `(hidden_size, num_output_tokens, 1)`, the `num_rows_in_final_output` should be set to `num_output_tokens` (not `hidden_size`) because of a swap+transpose operation that maps rows of the output tensor to `hidden_size` and columns to `num_output_tokens`.Learnt from: ixlmar Repo: NVIDIA/TensorRT-LLM PR: 7294 File: tensorrt_llm/_torch/pyexecutor/sampler.py:887-891 Timestamp: 2025-08-28T10:25:22.370Z Learning: In tensorrt_llm/_torch/pyexecutor/sampler.py, the draft_probs and target_probs tensors have shapes [1, steps] not [steps, vocab_size] as might be expected, making the .squeeze(0) operations appropriate for removing the batch dimension of size 1.
Collaborator
Author
|
/bot run --disable-fail-fast |
Collaborator
|
PR_Github #25173 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #25173 [ run ] completed with state |
Collaborator
Author
|
/bot run --disable-fail-fast |
Collaborator
|
PR_Github #25277 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #25277 [ run ] completed with state |
Collaborator
Author
|
/bot run --disable-fail-fast |
Collaborator
|
PR_Github #25323 [ run ] triggered by Bot. Commit: |
lfr-0531
reviewed
Nov 21, 2025
Collaborator
|
PR_Github #25323 [ run ] completed with state |
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #25365 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #25365 [ run ] completed with state |
heyuhhh
force-pushed
the
user/yuhangh/optimize_rocketkv_further
branch
from
519b120 to
7f77a2a
Compare
Collaborator
Author
|
/bot run --disable-fail-fast |
Collaborator
|
PR_Github #25559 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #25559 [ run ] completed with state |
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #25637 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #25637 [ run ] completed with state |
Collaborator
|
/bot run |
Collaborator
|
PR_Github #26088 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26088 [ run ] completed with state |
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #26145 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26145 [ run ] completed with state |
Collaborator
|
/bot run |
Collaborator
|
PR_Github #26186 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26186 [ run ] completed with state |
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #26215 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26215 [ run ] completed with state |
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #26234 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26234 [ run ] completed with state |
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #26245 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26245 [ run ] completed with state |
Collaborator
Author
|
/bot run |
Collaborator
|
PR_Github #26263 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26263 [ run ] completed with state |
Collaborator
|
/bot run |
Collaborator
|
PR_Github #26320 [ run ] triggered by Bot. Commit: |
Collaborator
|
PR_Github #26320 [ run ] completed with state |
Superjomn
pushed a commit
to hchings/TensorRT-LLM
that referenced
this pull request
Dec 1, 2025
Signed-off-by: yuhangh <58161490+heyuhhh@users.noreply.github.com>
MinaHuai
pushed a commit
to davidmlw/TensorRT-LLM
that referenced
this pull request
Dec 10, 2025
…VIDIA#8779) The performance results of some kernels could be easily affected by the warm/cold L2 cache status. To achieve more precise profiling results, the L2 cache is cleared for every execution by the circular buffer method for better benchmarking during autotuning. Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com> [None][infra] Waive failed cases for main branch on 11/25 (NVIDIA#9429) Signed-off-by: qqiao <qqiao@nvidia.com> [NVIDIA#8391][chore] test_perf.py to lock clocks read from gpu_configs.yml instead of max freq (NVIDIA#9409) Signed-off-by: Eran Geva <19514940+MrGeva@users.noreply.github.com> [None][ci] Move more test stages to use OCI machines (NVIDIA#9395) Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> Co-authored-by: Matt Lefebvre <matthewelefebvre@gmail.com> [None][feat] Improve TRTLLM MoE in small hidden size throughput cases (NVIDIA#9377) Signed-off-by: Anthony Chang <27950904+rosenrodt@users.noreply.github.com> [https://nvbugs/5537996][fix] Let KV cache manager block initialization be aware whether it is doing a dry run or not (NVIDIA#9093) Before this commit, the kv cache manager does the same regardless, which causes a mis-calculation in free memory available to allocate for the KV cache manager, hence causing a crash. This commit fixes this by letting KV cache manager initialization be aware whether it is doing the dry run or not. If it is a dry run, use the max_tokens setting that is already pre-calculated and filled into kv_cache_config.max_tokens. Signed-off-by: eopXD <yuehtingc@nvidia.com> [https://nvbugs/5667922][fix] Update long context evaluation config (NVIDIA#9426) Signed-off-by: mni <125171826+baize97@users.noreply.github.com> [None][fix] Mitigate test timeout issues (NVIDIA#9445) Signed-off-by: Shixiaowei02 <39303645+Shixiaowei02@users.noreply.github.com> [None][chore] Fix trtllm-eval for PyTorchLLM (NVIDIA#9427) Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com> [None][feat] Add a parser to layer-wise benchmarks (NVIDIA#9440) Signed-off-by: Tailing Yuan <yuantailing@gmail.com> [None][feat] Support custom chat template for tool calling (NVIDIA#9297) Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com> [TRTLLM-8160][feat] Add draft token tree runtime on CDL (NVIDIA#8586) Signed-off-by: Yue Weng <25103990+yweng0828@users.noreply.github.com> [None][ci] waive a test (NVIDIA#9458) Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com> [https://nvbugs/5680905][fix] Relax the MMLU accuracy requirement for DS-v3.2 (NVIDIA#9439) Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com> [TRTLLM-8376][feat] top-p optimization (removes redundant softmax) (NVIDIA#9411) Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com> [TRTLLM-9490][feat] use FlashInfer's top_k_sampling_from_probs (NVIDIA#9457) Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com> [https://nvbugs/5647400] [fix] Enlarged the AllReduce workspace size to 64MB. Added AllReduce strategy to AD config. (NVIDIA#9145) Signed-off-by: Eran Geva <19514940+MrGeva@users.noreply.github.com> [TRTLLM-909][feat] Overlap context chunks in pipeline parallel mode (NVIDIA#9308) Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> [None][chore] AutoDeploy add multi stream moe pass to default.yaml (NVIDIA#9430) Signed-off-by: Suyog Gupta <41447211+suyoggupta@users.noreply.github.com> [https://nvbugs/5685143][fix] avoid cudaFree overlap with cuda graph (NVIDIA#9438) Signed-off-by: Chuang Zhu <111838961+chuangz0@users.noreply.github.com> [None][chore] Bump version to 1.2.0rc5 (NVIDIA#9455) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> [TRTLLM-8936][test] Add disagg and wideep multi-node multi-gpu test cases (NVIDIA#9356) Signed-off-by: FredricZ-2007 <226039983+fredricz-20070104@users.noreply.github.com> [None][ci] move some slow test cases of DGX-B200 to post merge (NVIDIA#9467) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> [TRTLLM-9293][feat] Enable partial weight loading to support streaming update weights (NVIDIA#9224) Signed-off-by: shuyix <219646547+shuyixiong@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [TRTLLM-9264][fix] Add accuracy/unit tests/doc for phi4mm (NVIDIA#9246) Signed-off-by: Wanli Jiang <35160485+Wanli-Jiang@users.noreply.github.com> [https://nvbugs/5580099][fix] Cherry pick IMA issue fix from release/1.1 (NVIDIA#9032) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [None][chore] Upgrade CuteDSL to 4.3.0 (NVIDIA#9444) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [None][feat] Support MLA chunked prefill for DeepSeek V3.2 model (NVIDIA#9376) Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com> [None][feat] Add environment variable to force spec-dec number of accepted tokens (NVIDIA#9371) Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com> [None][infra] Update allowed list 2025.11.25 (NVIDIA#9468) Signed-off-by: Yuanjing Xue <197832395+yuanjingx87@users.noreply.github.com> [None][infra] Fail the pipeline when slurm ssh dropped (NVIDIA#9157) Signed-off-by: Yuanjing Xue <197832395+yuanjingx87@users.noreply.github.com> [None][feat] AutoDeploy: Remove redundant copies in mamba layers (NVIDIA#9461) Signed-off-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com> Co-authored-by: Suyog Gupta <41447211+suyoggupta@users.noreply.github.com> [None][feat] AutoDeploy: Add A_log fusion for Mamba layers (NVIDIA#9422) Signed-off-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com> [None][ci] Waive blackwell test on spec gate. (NVIDIA#9502) Signed-off-by: Zheyu Fu <zheyuf@NVIDIA.com> [https://nvbugs/5608930][fix] Fix a typo (NVIDIA#9487) Signed-off-by: Shixiaowei02 <39303645+Shixiaowei02@users.noreply.github.com> [NVIDIA#9463][feat] Add revision option to trtllm commands (NVIDIA#9498) Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com> [TRTLLM-9085][doc] fix math formula rendering issues (NVIDIA#9481) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> [None][chore] update comments in llm_args.py (NVIDIA#9472) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [https://nvbugs/5680310][fix] Fix ctx only timed out test (NVIDIA#9410) Signed-off-by: Patrice Castonguay <55748270+pcastonguay@users.noreply.github.com> [https://nvbugs/5547414][fix] enable case after using local cache model (NVIDIA#9473) Signed-off-by: Hui Gao <huig@nvidia.com> [None][fix] Replace PYTORCH_CUDA_ALLOC_CONF with PYTORCH_ALLOC_CONF to fix deprecation warning (NVIDIA#9294) Signed-off-by: Jiagan Cheng <jiaganc@nvidia.com> [https://nvbugs/5698581][fix] Init draft tokens for CUDA graph dummy request (NVIDIA#9505) Signed-off-by: ziyixiong-nv <219238287+ziyixiong-nv@users.noreply.github.com> [None][infra] Waive failed case in pre-merge on 11/27 (NVIDIA#9507) Signed-off-by: qqiao <qqiao@nvidia.com> [TRTLLM-9513][docs] Qwen3 deployment guide (NVIDIA#9488) Signed-off-by: Lanyu Liao <laliao@laliao-mlt.client.nvidia.com> Co-authored-by: Lanyu Liao <laliao@laliao-mlt.client.nvidia.com> [None][chore] revert batch_size=1 to prevent timeout and lower accuracy reference by 0.12% as a WAR (NVIDIA#9447) Signed-off-by: Lizhi Zhou <1432185+reasonsolo@users.noreply.github.com> Co-authored-by: Shi Xiaowei <39303645+Shixiaowei02@users.noreply.github.com> [TRTLLM-9279][infra] Use flexcache for gh200 nodes since they locate in Austin (NVIDIA#9405) Signed-off-by: qqiao <qqiao@nvidia.com> Signed-off-by: Emma Qiao <qqiao@nvidia.com> Co-authored-by: Yanchao Lu <yanchaol@nvidia.com> [cherry-pick][https://nvbugs/5670793][fix] Solve trtllm-serve launch_disaggregated issue (NVIDIA#9346) Signed-off-by: xxi <xxi@nvidia.com> [None][infra] Fix Slurm job script (NVIDIA#9508) Signed-off-by: Yuanjing Xue <197832395+yuanjingx87@users.noreply.github.com> [None][fix] change allreduce workspace dtype to torch.int64 to avoid overflow (NVIDIA#9479) Signed-off-by: Zhenhuan Chen <zhenhuanc@nvidia.com> [None][feat] add qwen3-next CI test of accuracy on BF16 and NVFP4 (NVIDIA#9330) Signed-off-by: jiant <107457950+JadoTu@users.noreply.github.com> [None][fix] fix TP support for DeepSeek-V3.2 on hopper (NVIDIA#9484) Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com> [TRTLLM-9389][chore] Refactor AlltoallMethodType. (NVIDIA#9388) Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com> [https://nvbugs/5674665][chore] Add test coverage for https://nvbugspro.nvidia.com/bug/5674665 (NVIDIA#9518) Signed-off-by: eopXD <yuehtingc@nvidia.com> [TRTLLM-7288][infra] Download merged waive list in slurm script (NVIDIA#8999) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> Co-authored-by: Yanchao Lu <yanchaol@nvidia.com> [https://nvbugs/5687820][fix] Remove self.abort() in DetokenizedGenerationResult (NVIDIA#9449) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [NVIDIA#9150][feat] AutoDeploy Nemotron-Flash support (NVIDIA#9504) Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com> [None] [chore] Update to cutlass 4.3 (NVIDIA#8637) Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> [https://nvbugs/5637037][chore] Update waive lists. (NVIDIA#9386) Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com> Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> Co-authored-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [TRTLLM-8970][infra] Fix generate report when has isolation test result (NVIDIA#8861) Signed-off-by: qqiao <qqiao@nvidia.com> Signed-off-by: Emma Qiao <qqiao@nvidia.com> [https://nvbugs/5685015][fix] Update invalid max_token test (NVIDIA#9435) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [None][fix] Fix on-disk cache and revise logger/statistics for AutoTuner. (NVIDIA#9211) Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com> [https://nvbugs/5689658][test] Fix gpu lock issue running on cluster (NVIDIA#9441) Signed-off-by: yufeiwu <230315618+yufeiwu-nv@users.noreply.github.com> [None][chore] add spec_decoding configs in perf benchmark scripts and fix typos (NVIDIA#9533) Signed-off-by: Lanyu Liao <lancelly@users.noreply.github.com> Co-authored-by: Lanyu Liao <lancelly@users.noreply.github.com> [None][fix] Remove FP8 K/V buffer from TRTLLM sparse MLA attention kernel (NVIDIA#9529) Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com> [None] [chore] Enhancements and clean up to slurm scripts (NVIDIA#9493) Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> [None][chore] Revert "[None][fix] change allreduce workspace dtype to torch.int64 t… (NVIDIA#9538) Signed-off-by: Zhenhuan Chen <zhenhuanc@nvidia.com> [None][infra] Waive failed cases for main branch on 11/28 (NVIDIA#9539) Signed-off-by: qqiao <qqiao@nvidia.com> [None][fix] Pass checkpoint_format to create_input_processor (NVIDIA#9521) Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> [TRTLLM-9541][infra] Use artifactory mirror for download.pytorch.org (NVIDIA#9477) Signed-off-by: ZhanruiSunCh <184402041+ZhanruiSunCh@users.noreply.github.com> Signed-off-by: Zhanrui Sun <184402041+ZhanruiSunCh@users.noreply.github.com> Co-authored-by: Yanchao Lu <yanchaol@nvidia.com> [TRTLLM-9488][feat] add 'disable_flashinfer_sampling' config option (NVIDIA#9454) Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com> [None][infra] Waive failed case in pre-merge on 11/28 (NVIDIA#9537) Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com> [None][perf] Helix: improve all-to-all perf for large CP size (NVIDIA#9494) Signed-off-by: Matthias Jouanneaux <mjoux@nvidia.com> Signed-off-by: Zheyu Fu <zheyuf@NVIDIA.com> Co-authored-by: Zheyu Fu <zheyuf@nvidia.com> [None][feat] support for more accurate AR calculation (NVIDIA#9323) Signed-off-by: binghanc <176802681+binghanc@users.noreply.github.com> [TRTLLM-9488][fix] llmapi references (NVIDIA#9547) Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com> [NVIDIA#8948][feat] Support custom sharding config (NVIDIA#9143) Signed-off-by: greg-kwasniewski1 <213329731+greg-kwasniewski1@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [None][chore] Weekly mass integration of release/1.1 -- rebase (NVIDIA#9522) Signed-off-by: yunruis <205571022+yunruis@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> Signed-off-by: Wangshanshan <30051912+dominicshanshan@users.noreply.github.com> Signed-off-by: qgai <qgai@nvidia.com> Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com> Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> Signed-off-by: Simeng Liu <simengl@nvidia.com> Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com> Signed-off-by: Ivy Zhang <25222398+crazydemo@users.noreply.github.com> Signed-off-by: Vincent Zhang <vinczhang@nvidia.com> Signed-off-by: peaceh <103117813+peaceh-nv@users.noreply.github.com> Signed-off-by: Michal Guzek <mguzek@nvidia.com> Signed-off-by: Michal Guzek <moraxu@users.noreply.github.com> Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com> Signed-off-by: leslie-fang25 <leslief@nvidia.com> Signed-off-by: Shunkang <182541032+Shunkangz@users.noreply.github.co> Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> Co-authored-by: yunruis <205571022+yunruis@users.noreply.github.com> Co-authored-by: sunnyqgg <159101675+sunnyqgg@users.noreply.github.com> Co-authored-by: brb-nv <169953907+brb-nv@users.noreply.github.com> Co-authored-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com> Co-authored-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com> Co-authored-by: Simeng Liu <109828133+SimengLiu-nv@users.noreply.github.com> Co-authored-by: Guoming Zhang <137257613+nv-guomingz@users.noreply.github.com> Co-authored-by: Jin Li <59594262+liji-nv@users.noreply.github.com> Co-authored-by: Ivy Zhang <25222398+crazydemo@users.noreply.github.com> Co-authored-by: Vincent Zhang <vcheungyi@163.com> Co-authored-by: peaceh-nv <103117813+peaceh-nv@users.noreply.github.com> Co-authored-by: Michal Guzek <moraxu@users.noreply.github.com> Co-authored-by: Chang Liu <9713593+chang-l@users.noreply.github.com> Co-authored-by: Leslie Fang <leslief@nvidia.com> Co-authored-by: Shunkangz <182541032+Shunkangz@users.noreply.github.com> Co-authored-by: Shunkang <182541032+Shunkangz@users.noreply.github.co> Co-authored-by: QI JUN <22017000+QiJune@users.noreply.github.com> [TRTLLM-5971][feat] Integrate helix parallelism (NVIDIA#9342) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [None][infra] - Request idle time exemption for OCI jobs (NVIDIA#9528) Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> [None][infra] Wiave failed tests for main branch on 11/30 (NVIDIA#9555) Signed-off-by: qqiao <qqiao@nvidia.com> [None][fix] Fix port conflict in disagg tests (NVIDIA#9474) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [None][ci] Split H100_PCIe-PyTorch-Post-Merge test stage (NVIDIA#9558) Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> [None][ci] Split H100_PCIe-PyTorch-Post-Merge test stage (NVIDIA#9559) Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> [TRTLLM-8958][feat] and [TRTLLM-8960]: create ConfigurableMoE and support TRTLLMGenFusedMoE as backend (NVIDIA#9486) [None] [feat] Optimize the algorithm part of RocketKV (NVIDIA#9333) Signed-off-by: yuhangh <58161490+heyuhhh@users.noreply.github.com> [https://nvbugs/5690172][fix] Fix Qwen3-235B ATP accuracy issue with PDL (NVIDIA#9530) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [TRTLLM-6222][feat] Extend cute_dsl_nvfp4_gemm to sm103. (NVIDIA#9543) Signed-off-by: Mindy Li <11663212+limin2021@users.noreply.github.com> [None][fix] Correct virtual memory allocation alignment (NVIDIA#9491) Signed-off-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [https://nvbugs/5684703][fix] Unwaive disagg guided decoding test (NVIDIA#9466) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [https://nvbugs/5503479][fix] Temporarily lower reference accuracy to stabilize CI (NVIDIA#9398) Signed-off-by: Pengbo Wang <221450789+pengbowang-nv@users.noreply.github.com> [None][chore] remove qwen3-next accuracy tests (NVIDIA#9534) Signed-off-by: jiant <107457950+JadoTu@users.noreply.github.com> [None][doc] fix mtp.py typo (NVIDIA#9307) Signed-off-by: liugaoji <757394026@qq.com> [None][feat] add chat template kwargs support to longbench-v2 (NVIDIA#9544) Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com> [NVIDIA#9496][fix] AutoDeploy: remove auto-tuner from nvfp4_gemm forward (NVIDIA#9497) Signed-off-by: Neta Zmora <96238833+nzmora-nvidia@users.noreply.github.com> [None][fix] Replace hash method with unique_id for cutedsl MoE runners. (NVIDIA#9569) Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com> [None][chore] refactor disaggregated scripts to use named arguments (NVIDIA#9581) Signed-off-by: Zhenhuan Chen <zhenhuanc@nvidia.com> [TRTLLM-6222][feat] Several perf opt for cuteDSL nvf4 gemm (NVIDIA#9428) Signed-off-by: Yuhan Li <51736452+liyuhannnnn@users.noreply.github.com> [None][chore] reduce the layers of the `devel` docker image (NVIDIA#9077) Signed-off-by: Martin Marciniszyn Mehringer <11665257+MartinMarciniszyn@users.noreply.github.com> [https://nvbugs/5651854][infra] Enable perf metrics during accuracy testing (NVIDIA#9140) [None][fix] Skip Allreduce init for Attention DP (NVIDIA#9542) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [None][test] [None][test] Waive main branch test failures 12/1 (NVIDIA#9566) Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> [None][ci] Minor change for Slurm scripts (NVIDIA#9561) Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> [TRTLLM-6768][infra] Fix params for not updating github status (NVIDIA#6747) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> [None][infra] Update the pytest options after MI (NVIDIA#9579) Signed-off-by: qqiao <qqiao@nvidia.com> [TRTLLM-6756][feat] Add Beam Search to TorchSampler (NVIDIA#8509) Signed-off-by: Stefan Niebler <82932102+stnie@users.noreply.github.com> [None][chore] Defer exposing context parallel configs (NVIDIA#9552) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> [TRTC-1943][feat] Env vars override support in LLM API (NVIDIA#9104) Signed-off-by: Venky Ganesh <23023424+venkywonka@users.noreply.github.com> [None][feat] AutoDeploy: Use the router gemm op for nemotron MOE (NVIDIA#9500) Signed-off-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com> [NVIDIA#9198][feat] Refactor dist ops in AutoDeploy (NVIDIA#9301) Signed-off-by: Eran Geva <19514940+MrGeva@users.noreply.github.com> [None][fix] Prevent YAML partial kv_cache_config from incorrectly overriding the complete kv_cache_config (NVIDIA#9262) Signed-off-by: Yuening Li <62227368+Yuening-wa@users.noreply.github.com> [TRTLLM-9085][doc] fix math formula rendering issues in github (NVIDIA#9605) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> [None][feat] Unify nvfp4 gemm backend (NVIDIA#8963) Signed-off-by: Shijie Wang <jaywan@nvidia.com> Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com> Signed-off-by: Shijie <jaywan@nvidia.com> Co-authored-by: Yukun He <23156053+hyukn@users.noreply.github.com> [None][feat] Add support for KVCache reuse for DSv32 (NVIDIA#9383) Signed-off-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [None][chroe] Polish qwen3-next modeling code. (NVIDIA#8902) Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> [https://nvbugs/5703953][fix] Use random port for disagg tests (NVIDIA#9582) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [None][fix] Waive gb200 (NVIDIA#9580) Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com> [FMDL-1328][feat] Add support for nano-v3 and super-v3 with pytorch backend (NVIDIA#9261) Signed-off-by: Wanli Jiang <35160485+Wanli-Jiang@users.noreply.github.com> [https://nvbugs/5582091][test] increase warmup times in testing for multi-gpu cases (NVIDIA#9578) Signed-off-by: Ruodi Lu <ruodil@users.noreply.github.com> Co-authored-by: Ruodi Lu <ruodil@users.noreply.github.com> [None][chore] Add failed cases into waives.txt (NVIDIA#9588) Signed-off-by: xinhe-nv <200704525+xinhe-nv@users.noreply.github.com> [https://nvbugs/5702793][fix] Fix uncontiguous tensor view (NVIDIA#9576) Signed-off-by: shuyix <219646547+shuyixiong@users.noreply.github.com> [None][infra] Waive failed cases for main branch (NVIDIA#9615) Signed-off-by: qqiao <qqiao@nvidia.com> [TRTLLM-9488][feat] use FlashInfer.sampling by default (NVIDIA#9545) Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com> [None][infra] Update allowlist 2025/12/01 (NVIDIA#9616) Signed-off-by: Yuanjing Xue <197832395+yuanjingx87@users.noreply.github.com> [None][infra] Remove an invalid test name in waives.txt (NVIDIA#9620) Signed-off-by: qqiao <qqiao@nvidia.com> Lock the gpu clocks in L0 perf tests (NVIDIA#9585) Signed-off-by: Eran Geva <19514940+MrGeva@users.noreply.github.com> [TRTLLM-9466][test] Evaluate helix parallelism with DSV3 Lite (NVIDIA#9597) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> [None][fix] Extract GPU count from single-node stage names (NVIDIA#9599) Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com> [https://nvbugs/5667774][fix] Refine Piecewise Cuda Graph Condition for DP (NVIDIA#9393) Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com> [TRTLLM-9144][fix] enhance RPC robustness (NVIDIA#8711) Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com> Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com> Co-authored-by: Erin Ho <14718778+hchings@users.noreply.github.com> [https://nvbugs/5627710][fix] Fix synchronization bugs in KvCacheTransferManager that can cause corrupted blocks (NVIDIA#9056) Signed-off-by: thorjohnsen <41591019+thorjohnsen@users.noreply.github.com> Signed-off-by: Thor Johnsen <41591019+thorjohnsen@users.noreply.github.com> Co-authored-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com> Co-authored-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> [TRTLLM-8980][test] Clean up spec dec tests in test_llm_api_pytorch (NVIDIA#8889) Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [NVIDIA#9150][feat] Add code for nano v3 to custom implementation in AD (NVIDIA#9465) * Why? We would like to show an alternative to monkey-patching in AutoDeploy. * What? This commit builds on the existing custom model implementation for NemotronH and adds the bits relevant for MoE layers. Part of NVIDIA#9150. Signed-off-by: William Zhang <133824995+2ez4bz@users.noreply.github.com> [NVIDIA#9150][feat] AutoDeploy: reviewer comments for NVIDIA#9150 (NVIDIA#9527) Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com> [https://nvbugs/5651854][fix] Fix dist-serving perf by clearing CPU affinity (NVIDIA#9549) Signed-off-by: Shixiaowei02 <39303645+Shixiaowei02@users.noreply.github.com> [NVIDIA#9550][feat] AutoDeploy: Add NVFP4 Cutlass MoE kernels (NVIDIA#9551) Signed-off-by: Neta Zmora <96238833+nzmora-nvidia@users.noreply.github.com> [https://nvbugs/5688388][fix] fix: Reducing num request in disagg test to speed up (NVIDIA#9598) Signed-off-by: Patrice Castonguay <55748270+pcastonguay@users.noreply.github.com> [TRTLLM-8946][feat] Improved heuristics to detect shardable regions (NVIDIA#9200) Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com> Signed-off-by: greg-kwasniewski1 <213329731+greg-kwasniewski1@users.noreply.github.com> Co-authored-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com> [NVIDIA#9632][feat] Support EXTRA_WHEEL_BUILD_ARGS during wheel build (NVIDIA#9633) Signed-off-by: Yu Chi Li <yuchil@nvidia.com> [None][chore] Waive test failing on pre-merge (NVIDIA#9638) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> [None][chore] Remove traceback dump for multimodal input processor (NVIDIA#9634) Signed-off-by: Chang Liu (Enterprise Products) <9713593+chang-l@users.noreply.github.com> [None][chore] Fix trtllm-eval and move GroupedGemmInputsHelper (NVIDIA#9612) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [https://nvbugs/5698434][fix] Use separate weight mapper for draft (NVIDIA#9607) Signed-off-by: Anurag Mukkara <134339030+amukkara@users.noreply.github.com> [TRTLLM-7101][infra] Reuse passed tests (NVIDIA#6894) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> Co-authored-by: Yanchao Lu <yanchaol@nvidia.com> [None][test] Remove duplicate test cases (NVIDIA#9623) Signed-off-by: yufeiwu <230315618+yufeiwu-nv@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [None][feat] Add RocketKV usage doc and e2e accuracy test on LongBenchV2 (NVIDIA#9572) Signed-off-by: yuhangh <58161490+heyuhhh@users.noreply.github.com> [TRTLLM-9242][doc] Add examples showcasing openai compatible APIs (NVIDIA#9520) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [None][chore] AutoDeploy update cuda stream manager for multi-device (NVIDIA#9575) Signed-off-by: Suyog Gupta <41447211+suyoggupta@users.noreply.github.com> [TRTLLM-9391][chore] Automatically estimate required workspace. (NVIDIA#9535) Signed-off-by: Bo Li <22713281+bobboli@users.noreply.github.com> [https://nvbugs/5708475][fix] Fix e2e eval accuracy for helix parallelism (NVIDIA#9647) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> [https://nvbugs/5561153][test] Fix log error for perf test (NVIDIA#9622) Signed-off-by: FredricZ-2007 <226039983+fredricz-20070104@users.noreply.github.com> [TRTLLM-8241][feat] Aliasing to comply to LlmArgs (NVIDIA#9586) Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com> [None][chore] Add failed cases into waives.txt (NVIDIA#9593) Signed-off-by: Jie Li <lijie@nvidia.com> Co-authored-by: Jie Li <lijie@nvidia.com> [TRTLLM-6842][feat] Support Response API for general purpose (NVIDIA#9392) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [None][test] Update Qwen3-next accuracy testing by setting the cuda … (NVIDIA#9613) Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> [None][feat] update trtllm-gen nvfp4 kernels with better performance (NVIDIA#9510) Signed-off-by: Perkz Zheng <67892460+PerkzZheng@users.noreply.github.com> [None][doc] Replace the tensorrt icon with torch icon on overview.md (NVIDIA#9644) Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> [https://nvbugs/5705197][chore] Unwaive timeout disagg tests (NVIDIA#9637) Signed-off-by: Patrice Castonguay <55748270+pcastonguay@users.noreply.github.com> [https://nvbugs/5552132][fix] Enable LoRa for GPT OSS Torch (NVIDIA#8253) Signed-off-by: Michal Guzek <mguzek@nvidia.com> [None][fix] Fix wide ep MoE error (NVIDIA#9642) Signed-off-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com> [https://nvbugs/5702795][fix] Remove the warning message for aten.log. (NVIDIA#9665) Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> [https://nvbugs/5693853][fix] Fix error handling when querying machin… (NVIDIA#9483) Signed-off-by: Gal Hubara Agam <96368689+galagam@users.noreply.github.com> [OMNIML-2932] [feat] nvfp4 awq support (NVIDIA#8698) Signed-off-by: weimingc <17592131+meenchen@users.noreply.github.com> [NVIDIA#9643][fix] AutoDeploy: fix nano sharding config (NVIDIA#9668) Signed-off-by: Lucas Liebenwein <11156568+lucaslie@users.noreply.github.com> [NVIDIA#9147][feat] AutoDeploy: Draft Target Speculative Decoding (NVIDIA#9275) Signed-off-by: Govind Ramnarayan <105831528+govind-ramnarayan@users.noreply.github.com> [None][feat] Update Qwen3CodeToolParser to align tool-calling parameters (NVIDIA#9540) Signed-off-by: Wanli Jiang <35160485+Wanli-Jiang@users.noreply.github.com> [TRTLLM-7181][infra] Generate test results when pytest timeout happens (NVIDIA#9396) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [TRTLLM-9522][fix] restore `trtllm-serve mm_embedding_serve` (NVIDIA#9669) [TRTLLM-5093][infra] Write env variables to a file in the interactive debug session (NVIDIA#6792) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> [None][fix] fix error when processing batches containing both text and mm data (NVIDIA#8381) Signed-off-by: Nekofish-L <liuxiangyang@mail.ustc.edu.cn> [TRTLLM-7073][feat] Support torch compile for PP for Llama and DeepSeekV3 (NVIDIA#7838) Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com> [None][feat] Add weights initialization and context phase parser to layer-wise benchmarks (NVIDIA#9667) Signed-off-by: Tailing Yuan <yuantailing@gmail.com> [TRTLLM-8274][feat] Check if executor is shutdown in /health entrypoint (NVIDIA#9057) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [NVIDIA#8733][feat] Add Llama4 MoE handling to AutoDeploy (NVIDIA#9556) Signed-off-by: Tal Cherckez <127761168+tcherckez-nvidia@users.noreply.github.com> Signed-off-by: tcherckez-nvidia <127761168+tcherckez-nvidia@users.noreply.github.com> Co-authored-by: Neta Zmora <nzmora@nvidia.com> [None][ci] unwaive tests (NVIDIA#9651) Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com> [None][feat] Add NIXL-LIBFABRIC support (NVIDIA#9225) Signed-off-by: Yoray Zack <62789610+zackyoray@users.noreply.github.com> Signed-off-by: zackyoray <yorayz@nvidia.com> [None][test] rename wide ep and disagg metric name in perf test (NVIDIA#9704) Signed-off-by: Ruodi Lu <ruodil@users.noreply.github.com> Co-authored-by: Ruodi Lu <ruodil@users.noreply.github.com> [https://nvbugs/5467531][fix] Unwaive fused_moe all to all test with … (NVIDIA#9617) Signed-off-by: Jin Li <59594262+liji-nv@users.noreply.github.com> [None][fix] Recover TRTLLM MoE Perf for DEP (NVIDIA#9562) Signed-off-by: Anthony Chang <27950904+rosenrodt@users.noreply.github.com> [None][chore] Add failed cases into waives.txt (NVIDIA#9662) Signed-off-by: Xin He (SW-GPU) <200704525+xinhe-nv@users.noreply.github.com> Signed-off-by: xinhe-nv <200704525+xinhe-nv@users.noreply.github.com> Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> Co-authored-by: Yanchao Lu <yanchaol@nvidia.com> [None][fix] Fix TLLM_SPEC_DECODE_FORCE_NUM_ACCEPTED_TOKENS for MTP/EAGLE (NVIDIA#9608) Signed-off-by: Aurelien Chartier <2567591+achartier@users.noreply.github.com> [None][infra] Add container notices and documentation (NVIDIA#9185) Signed-off-by: Parker Drake <pdrake@nvidia.com> [TRTLLM-5312][infra] Add triton trigger rules (NVIDIA#6440) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> [None][doc] Add feature docs for helix parallelism (NVIDIA#9684) Signed-off-by: Balaram Buddharaju <169953907+brb-nv@users.noreply.github.com> [TRTLLM-9579][infra] Set mergeWaiveList stage UNSTABLE when there is any issue (NVIDIA#9692) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> [None][doc] Added line about partial reuse (NVIDIA#7846) Signed-off-by: thorjohnsen <41591019+thorjohnsen@users.noreply.github.com> [TRTLLM-8920][feat] decouple disagg service from fastapi (NVIDIA#8714) Signed-off-by: Lizhi Zhou <1432185+reasonsolo@users.noreply.github.com> [https://nvbugs/5633340][fix] start disagg workers and servers on free ports (NVIDIA#9694) Signed-off-by: Lizhi Zhou <1432185+reasonsolo@users.noreply.github.com> [TRTLLM-9562] [doc] Add Deployment Guide for Kimi K2 Thinking on TensorRT LLM - Blackwell (NVIDIA#9711) Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> [NVIDIA#9602][feat] AutoDeploy: Support TRTLLM Sampler (NVIDIA#9641) Signed-off-by: Govind Ramnarayan <105831528+govind-ramnarayan@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [None] [tests] Unwaive EPLB tests (NVIDIA#9625) Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> [https://nvbugs/5518713][test] Refactor core test lists by merging with llm_perf_cluster.yml (NVIDIA#9714) Signed-off-by: yufeiwu <230315618+yufeiwu-nv@users.noreply.github.com> [TRTLLM-7136][feat] Update load_weights method to include mapping parameter in checkpoint loaders (NVIDIA#9583) Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> [None][refactor] Improve request processing function in sampler (NVIDIA#9671) Signed-off-by: Robin Kobus <19427718+Funatiq@users.noreply.github.com> [https://nvbugs/5670672][fix] Fix flaky KV connector tests (NVIDIA#9676) Signed-off-by: jthomson04 <jwillthomson19@gmail.com> [None][infra] Update allowed list 20251204 (NVIDIA#9718) Signed-off-by: Yuanjing Xue <197832395+yuanjingx87@users.noreply.github.com> [None][feat] AutoDeploy: Perf optimization for Attention and rmsnorm (NVIDIA#9719) Signed-off-by: Chenghao Zhang <211069071+nvchenghaoz@users.noreply.github.com> [None][chore] Waive flakey disagg tests (NVIDIA#9749) Signed-off-by: Mike Iovine <miovine@nvidia.com> [https://nvbugs/5601682][fix] Fix cacheTransceiver hang (NVIDIA#9311) Signed-off-by: Iman Tabrizian <10105175+tabrizian@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [TRTLLM-9199][docs] KV Connector Docs (NVIDIA#9325) Signed-off-by: jthomson04 <jwillthomson19@gmail.com> Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [TRTLLM-9160][doc] add doc to llm_runtime.py (NVIDIA#9482) Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [None][doc] VDR 1.0 trtllm-serve doc enhancement (NVIDIA#9443) Signed-off-by: Pengyun Lin <81065165+LinPoly@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [TRTLLM-9086][doc] Clean up TODOs in documentation (NVIDIA#9292) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [TRTLLM-9157][doc] Guided decoding doc improvement (NVIDIA#9359) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [None][infra] Updated Linux installation guide (NVIDIA#9485) Signed-off-by: Yiqing Yan <yiqingy@nvidia.com> Co-authored-by: Yanchao Lu <yanchaol@nvidia.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [TRTLLM-9075][doc] refine the slurm examples (NVIDIA#9548) Signed-off-by: Yan Chunwei <328693+Superjomn@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [TRTLLM-9093][doc] update hyper links in overview (NVIDIA#9568) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [TRTLLM-9092][doc] link to modelopt checkpoints in quick start guide (NVIDIA#9571) Signed-off-by: junq <22017000+QiJune@users.noreply.github.com> Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com> Signed-off-by: Mike Iovine <miovine@nvidia.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [None][fix] Fix triton moe load_weight (NVIDIA#9649) Signed-off-by: shuyix <219646547+shuyixiong@users.noreply.github.com> [None][fix] fix a bug: deepseek_fp8_block_scales in TRTLLMGEN-MoE use 2D x_sf instead of 1D (NVIDIA#9658) Signed-off-by: xxi <xxi@nvidia.com> [TRTLLM-9372][feat] Enable CuteDSL MoE with Large EP (NVIDIA#9592) Signed-off-by: Enwei Zhu <21126786+syuoni@users.noreply.github.com> [TRTLLM-9522][chore] implement default `attach_multimodal_embeddings` (NVIDIA#9664) Signed-off-by: ixlmar <206748156+ixlmar@users.noreply.github.com> [TRTLLM-9660][feat] Convert cuteDSL GEMM to opt-in feature (NVIDIA#9682) Signed-off-by: Jonas Li <6110159+longlee0622@users.noreply.github.com> Co-authored-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> [None][fix] enable hmac in RPC (NVIDIA#9745) Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [https://nvbugs/5703953][fix] Preserving ip:port for trtllm-serve before initializing llm (NVIDIA#9646) Signed-off-by: Junyi Xu <219237550+JunyiXu-nv@users.noreply.github.com> [None][infra] Waive failed cases for main branch on 12/07 (NVIDIA#9769) Signed-off-by: qqiao <qqiao@nvidia.com> [None][fix] Several minor fixes to CI setting (NVIDIA#9765) Signed-off-by: Yanchao Lu <yanchaol@nvidia.com> [OMNIML-3036][doc] Re-branding TensorRT-Model-Optimizer as Nvidia Model-Optimizer (NVIDIA#9679) Signed-off-by: Chenjie Luo <chenjiel@nvidia.com> [None][feat] Enable NCCL_SYMMETRIC as default fallback for AllReduce (NVIDIA#9314) Signed-off-by: Ludwig Schneider <lschneider@nvidia.com> [TRTLLM-9000][feat] Add multi-node Perf Tests into CI (NVIDIA#8800) Signed-off-by: Chenfei Zhang <chenfeiz@nvidia.com> [None][test] add ntp tolerance in time metrics verification (NVIDIA#9741) Signed-off-by: zhengd-nv <200704041+zhengd-nv@users.noreply.github.com> [TRTLLM-9603][feat] Enable ConfigurableMoE test in the CI (NVIDIA#9645) [https://nvbugs/5422621][test] Add GB 200 WIDEEP test case for RCCA 5422621 (NVIDIA#9506) Signed-off-by: FredricZ-2007 <226039983+fredricz-20070104@users.noreply.github.com> [None][fix] Fix two tuning cache miss issues. (NVIDIA#9743) Signed-off-by: Yukun He <23156053+hyukn@users.noreply.github.com> [None][infra] Check in most recent lock file from nightly pipeline Signed-off-by: TensorRT LLM <90828364+tensorrt-cicd@users.noreply.github.com> [TRTLLM-9706] [doc] Update wide EP documents (NVIDIA#9724) Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> [https://nvbugs/5666804][test] only adding sampler config for limited models (NVIDIA#9512) Signed-off-by: Ruodi Lu <ruodil@users.noreply.github.com> Co-authored-by: Ruodi Lu <ruodil@users.noreply.github.com> Co-authored-by: yufeiwu-nv <230315618+yufeiwu-nv@users.noreply.github.com> Co-authored-by: Larry Xu <197874197+LarryXFly@users.noreply.github.com> [None][infra] Waive failed cases for main on 12/08 (NVIDIA#9773) Signed-off-by: qqiao <qqiao@nvidia.com> [None][chore] Move the rocketkv e2e test to post-merge (NVIDIA#9768) Signed-off-by: Fanrong Li <23290157+lfr-0531@users.noreply.github.com> [None][chore] Enable tvm_ffi for cute dsl nvfp4_gemm to reduce host overhead. (NVIDIA#9690) Signed-off-by: Mindy Li <11663212+limin2021@users.noreply.github.com> [TRTLLM-9431][perf] Enable multistream for Linear Attention in Qwen3-… (NVIDIA#9696) Signed-off-by: nv-guomingz <137257613+nv-guomingz@users.noreply.github.com> [None][chore] Remove closed bugs (NVIDIA#9770) Signed-off-by: xinhe-nv <200704525+xinhe-nv@users.noreply.github.com> [None][infra] update mooncake in docker images (NVIDIA#9584) Signed-off-by: zhengd-nv <200704041+zhengd-nv@users.noreply.github.com> Signed-off-by: Zheng Duan <200704041+zhengd-nv@users.noreply.github.com> [None][test] Add Kimi k2 WIDEEP perf and accuracy cases (NVIDIA#9686) Signed-off-by: FredricZ-2007 <226039983+fredricz-20070104@users.noreply.github.com> Signed-off-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> Co-authored-by: Kaiyu Xie <26294424+kaiyux@users.noreply.github.com> [https://nvbugs/5527655][test] Add test case for RCCA 5527655 (NVIDIA#9511) Signed-off-by: FredricZ-2007 <226039983+fredricz-20070104@users.noreply.github.com> [http://nvbugs/5649010][fix] fix test_auto_scaling.py::test_worker_restart timeout (NVIDIA#9775) Signed-off-by: Lizhi Zhou <1432185+reasonsolo@users.noreply.github.com> [None][fix] Switch AutoDeploy's default allreduce strategy to NCCL (NVIDIA#9666) Signed-off-by: Eran Geva <19514940+MrGeva@users.noreply.github.com> [TRTLLM-9506][fix] Fix AR for DeepSeek-R1 2 model path (NVIDIA#9661) Signed-off-by: qgai <qgai@nvidia.com> ray + updatew works trtllm works in async env trtllm works in sync and async env ray + updatew works rebase to the updated verl server mode still cherry pick still cherry pick still cherry pick integrated http interface hang at RyExecutor create workers ray.remote clean code use tensorrt_llm.rlhf_utils Signed-off-by: Liwei Ma <liweim@nvidia.com> placement, asyncllm, and basic tests Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> connect sleep and wakeup; Add support to pass None to update_weights Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> Batching ctx for IFB scheduler Signed-off-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com> accuracy WAR for TP>1: always use AllReduceStrategy.NCCL, refactored Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> fix e2e integration Signed-off-by: Superjomn <328693+Superjomn@users.noreply.github.com> update asyncllm, other nits Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> fix init setup Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> Fix TRTLLMSampler logprobs perf Signed-off-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com> fix and cleanup Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> fix server Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com> Revert "Batching ctx for IFB scheduler" This reverts commit b51aac0 Signed-off-by: Yuan Tong <13075180+tongyuantongyu@users.noreply.github.com> update & address comments Signed-off-by: Erin Ho <14718778+hchings@users.noreply.github.com>
codego7250
pushed a commit
to codego7250/TensorRT-LLM
that referenced
this pull request
Dec 11, 2025
Signed-off-by: yuhangh <58161490+heyuhhh@users.noreply.github.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Summary by CodeRabbit
New Features
--topk(replaces--topr),--kt_cache_dtype, and--tokens_per_blockfor enhanced sparse attention configuration.Updates
tokens_per_blockparameter for improved KV cache control.✏️ Tip: You can customize this high-level summary in your review settings.
Description
In this PR, mainly optimized the algorithm part of RocketKV based on the first triton implements (PR8682). The main optimizations are:
After the optimization, using RocketKV can achieve higher throughput and lower latency than before:
The accuracy is also competitive with TRTLLM baseline on Longbench v1 task:
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 thestage-listparameter 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.mdand the
scripts/test_to_stage_mapping.pyhelper.kill
killKill all running builds associated with pull request.
skip
skip --comment COMMENTSkip 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-pipelineReuse 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.