kanshan/TensorRT-LLMs
1
0
Fork 0
mirror of https://github.com/NVIDIA/TensorRT-LLM.git synced 2026-01-13 22:18:36 +08:00
Code Issues Actions 1 Packages Projects Releases Wiki Activity

[https://nvbugs/5522462][fix] Fix FP8 scout illegal memory access (#7845)

Browse Source 
Signed-off-by: Mike Iovine <6158008+mikeiovine@users.noreply.github.com>
This commit is contained in:
Mike Iovine 2025-09-19 10:30:37 -04:00 committed by GitHub
parent fbe325ce57
commit 8030b540ac
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 13 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
cpp/tensorrt_llm/kernels/decoderMaskedMultiheadAttention/decoderMaskedMultiheadAttentionTemplate.h
View File

@ -1363,7 +1363,8 @@ __global__ void __launch_bounds__(MAX_THEADS_PER_BLOCK, MIN_BLOCKS_PER_SM) maske
#ifndef MMHA_USE_FP32_ACCUM_FOR_LOGITS
if (sizeof(Tk) != 4)
{
auto const max_timesteps = min(timestep, static_cast<unsigned>(cyclic_kv_cache_len));
auto const max_timesteps
= min(timestep, min(chunked_attention_size, static_cast<unsigned>(cyclic_kv_cache_len)));
logits_smem_ += divUp(max_timesteps + 1, 4u) * 16;
}
Tk* logits_smem = reinterpret_cast<Tk*>(logits_smem_);

11
tests/integration/test_lists/test-db/l0_dgx_b200.yml
View File

@ -39,7 +39,18 @@ l0_dgx_b200:
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[tep4_latency_moe_trtllm-torch_compile=True]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_trtllm-torch_compile=False]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=False]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp8-cuda_graph=False]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp8ep4-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp8ep8-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp4-cuda_graph=False]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp4ep2-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_auto_dtype[tp4ep4-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp8[tp8ep8-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp8[tp4-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp4[tp8ep8-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp4[tp4-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp8_chunked_prefill[tp4ep4-cuda_graph=True]
- accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp4_chunked_prefill[tp4ep4-cuda_graph=True]
- disaggregated/test_disaggregated.py::test_disaggregated_deepseek_v3_lite_fp8_ucx[DeepSeek-V3-Lite-fp8]
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[tp4-trtllm-auto]
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[ep4-cutlass-auto]

1
tests/integration/test_lists/waives.txt
View File

@ -334,7 +334,6 @@ full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8
full:H100/accuracy/test_llm_api_pytorch.py::TestLlama4MaverickInstruct::test_fp8_eagle3[tp8-torch_compile=True] SKIP (https://nvbugs/5483534)
full:A100/test_e2e.py::test_ptp_quickstart_multimodal[NVILA-8B-FP16-vila/NVILA-8B-video-False] SKIP (https://nvbugs/5453725)
test_e2e.py::test_ptp_scaffolding[DeepSeek-R1-Distill-Qwen-7B-DeepSeek-R1/DeepSeek-R1-Distill-Qwen-7B] SKIP (https://nvbugs/5517260)
accuracy/test_llm_api_pytorch.py::TestLlama4ScoutInstruct::test_fp8_chunked_prefill[tp4ep4-cuda_graph=True] SKIP (https://nvbugs/5522462)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_bfloat16_4gpus[ep4-mtp_nextn=2-attention_dp=True-cuda_graph=True-overlap_scheduler=True-torch_compile=False] SKIP (https://nvbugs/5522746)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_nvfp4_4gpus[moe_backend=CUTLASS-mtp_nextn=0-tp2pp2-fp8kv=False-attention_dp=False-cuda_graph=False-overlap_scheduler=False-torch_compile=False] SKIP (https://nvbugs/5522746)
test_e2e.py::test_ptp_quickstart_multimodal[NVILA-8B-FP16-vila/NVILA-8B-image-False] SKIP (https://nvbugs/5523925)