Skip to content

[pull] main from NVIDIA:main #44

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
pull merged 10 commits into from
Jan 6, 2026
Merged

[pull] main from NVIDIA:main #44

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
1b10582
[TRTLLM-8638][fix] Add failed cases into waives.txt (#10384)
xinhe-nv Jan 6, 2026
22a1d31
[None][test] update test case constraint (#10381)
crazydemo Jan 6, 2026
810249c
[https://nvbugs/5769926] [fix] Add no container mount home WAR (#10431)
kaiyux Jan 6, 2026
9985277
[TRTLLM-8638][fix] Add failed cases into waives.txt (#10367)
xinhe-nv Jan 6, 2026
5108a69
[TRTLLM-9622][infra] Enable DGX_B300 multi-gpu testing in pre-merge p…
yiqingy0 Jan 6, 2026
1e82858
[TRTLLM-9896][test] add vswa test cases coverage (#10146)
crazydemo Jan 6, 2026
2eaabd7
[None] [fix] Fix undefined tokens_per_block (#10438)
kaiyux Jan 6, 2026
ab58d7c
[https://nvbugs/5772361][ci] Unwaive tests that have been fixed (#10424)
2ez4bz Jan 6, 2026
df0b976
[https://nvbugs/5785206][infra] Waive TestQwen3_30B_A3B::test_fp8[lat…
bobboli Jan 6, 2026
6507087
[None][infra] Waive failed cases on 1/6 (#10440)
EmmaQiaoCh Jan 6, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions examples/disaggregated/slurm/benchmark/submit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -287,7 +287,7 @@ def submit_job(config, log_dir, dry_run):
f"--container-image {env_config['container_image']}",
f"--container-name {container_name}",
f"--container-mounts {env_config['container_mount']}",
"--mpi=pmix --overlap",
"--no-container-mount-home --mpi=pmix --overlap",
f"bash {os.path.join(env_config['work_dir'], 'start_worker.sh')}",
server_type,
str(server_id),
Expand All @@ -313,7 +313,7 @@ def submit_job(config, log_dir, dry_run):
f"--container-name={container_name}",
f"--container-image={env_config['container_image']}",
f"--container-mounts={env_config['container_mount']}",
f"--mpi=pmix --overlap -N 1 -n 1",
f"--no-container-mount-home --mpi=pmix --overlap -N 1 -n 1",
f"bash {env_config['work_dir']}/start_server.sh {os.path.join(log_dir, 'server_config.yaml')} \"{server_env_var}\"",
f"&> {log_dir}/4_output_server.log &",
]
Expand Down
3 changes: 2 additions & 1 deletion jenkins/L0_Test.groovy
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3256,12 +3256,13 @@ def launchTestJobs(pipeline, testFilter)
"DGX_H100-4_GPUs-PyTorch-GptOss-1": ["dgx-h100-x4-oci", "l0_dgx_h100", 1, 1, 4],
"DGX_H100-4_GPUs-PyTorch-Others-1": ["dgx-h100-x4-oci", "l0_dgx_h100", 1, 1, 4],
"DGX_H100-4_GPUs-PyTorch-Ray-1": ["dgx-h100-x4-oci", "l0_dgx_h100", 1, 1, 4],
"B300-PyTorch-1": ["b300-single", "l0_b300", 1, 1],
"DGX_B200-4_GPUs-PyTorch-1": ["b200-x4-lbd", "l0_dgx_b200", 1, 1, 4, 1, true],
"DGX_B200-4_GPUs-PyTorch-Ray-1": ["b200-x4-lbd", "l0_dgx_b200", 1, 1, 4, 1, true],
"DGX_B200-8_GPUs-PyTorch-1": ["b200-x8-lbd", "l0_dgx_b200", 1, 1, 8, 1, true],
"DGX_B200-4_GPUs-PyTorch-Post-Merge-1": ["b200-x4-lbd", "l0_dgx_b200", 1, 2, 4, 1, true],
"DGX_B200-4_GPUs-PyTorch-Post-Merge-2": ["b200-x4-lbd", "l0_dgx_b200", 2, 2, 4, 1, true],
"B300-PyTorch-1": ["b300-single", "l0_b300", 1, 1],
"DGX_B300-4_GPUs-PyTorch-1": ["b300-x4", "l0_dgx_b300", 1, 1, 4],
"DGX_B300-4_GPUs-PyTorch-Post-Merge-1": ["b300-x4", "l0_dgx_b300", 1, 2, 4],
"DGX_B300-4_GPUs-PyTorch-Post-Merge-2": ["b300-x4", "l0_dgx_b300", 2, 2, 4],
// PerfSanity post-merge tests
Expand Down
5 changes: 3 additions & 2 deletions tensorrt_llm/_torch/attention_backend/sparse/dsa.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -813,13 +813,14 @@ def on_update_kv_lens(self):
# Expand schedule metadata buffer (only generation)
kv_lens_expanded = self.kv_lens_expanded_cuda[:num_tokens]
scheduler_metadata_buffer_expanded = get_paged_mqa_logits_metadata(
kv_lens_expanded, tokens_per_block, self.num_sms)
kv_lens_expanded, self.kv_cache_manager.tokens_per_block,
self.num_sms)
self.scheduler_metadata_buffer_expanded.copy_(
scheduler_metadata_buffer_expanded, non_blocking=True)
elif self.max_draft_tokens == 3:
scheduler_metadata_buffer_mtp3 = get_paged_mqa_logits_metadata(
self.kv_lens_cuda[self.num_contexts:self.num_seqs],
tokens_per_block, self.num_sms // 2)
self.kv_cache_manager.tokens_per_block, self.num_sms // 2)
self.scheduler_metadata_buffer_mtp3.copy_(
scheduler_metadata_buffer_mtp3, non_blocking=True)
self.prepare_dense_topk_indices(self.kv_lens_cuda, device=True)
Expand Down
11 changes: 11 additions & 0 deletions tests/integration/defs/accuracy/references/json_mode_eval.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,3 +8,14 @@ deepseek-ai/DeepSeek-V3-Lite:
- accuracy: 77.00
- spec_dec_algo: MTP
accuracy: 77.00
google/gemma-3-1b-it:
- quant_algo: FP8
kv_cache_quant_algo: FP8
accuracy: 61.00
GPT-OSS/120B-MXFP4:
- quant_algo: W4A16_MXFP4
spec_dec_algo: Eagle
accuracy: 62.00
- quant_algo: W4A8_MXFP4_MXFP8
spec_dec_algo: Eagle
accuracy: 62.00
140 changes: 140 additions & 0 deletions tests/integration/defs/accuracy/test_llm_api_pytorch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1105,6 +1105,37 @@ def test_fp8_prequantized(self):
task = MMLU(self.MODEL_NAME)
task.evaluate(llm)

def test_fp8_vswa_reuse(self):
# NOTE: Test with VSWA kv cache config.
kv_cache_config = KvCacheConfig(
enable_block_reuse=True,
max_attention_window=[512, 512, 512, 512, 512, 32768],
)
prequantized_model_path = f"{llm_models_root()}/gemma/gemma-3-1b-it-fp8/"
with LLM(prequantized_model_path,
kv_cache_config=kv_cache_config) as llm:
task = GSM8K(self.MODEL_NAME)
task.evaluate(llm)
task = MMLU(self.MODEL_NAME)
task.evaluate(llm)

@pytest.mark.parametrize("backend", ["xgrammar"])
def test_fp8_guided_decoding_vswa_reuse(self, backend: str, mocker):
mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
prequantized_model_path = f"{llm_models_root()}/gemma/gemma-3-1b-it-fp8/"
kv_cache_config = KvCacheConfig(
enable_block_reuse=True,
max_attention_window=[512, 512, 512, 512, 512, 32768],
)
cuda_graph_config = CudaGraphConfig(enable_padding=True)
llm = LLM(prequantized_model_path,
guided_decoding_backend=backend,
kv_cache_config=kv_cache_config,
cuda_graph_config=cuda_graph_config)
with llm:
task = JsonModeEval(self.MODEL_NAME)
task.evaluate(llm)

def test_auto_dtype_vswa_without_reuse(self):
# NOTE: Test with VSWA kv cache config.
kv_cache_config = KvCacheConfig(
Expand Down Expand Up @@ -2269,6 +2300,7 @@ def test_nvfp4_multi_gpus(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
torch.cuda.empty_cache()

@skip_pre_blackwell
@pytest.mark.skip_less_device_memory(95000)
@pytest.mark.parametrize(
"tp_size,pp_size,ep_size,mtp_nextn,fp8kv,attention_dp,cuda_graph,overlap_scheduler,max_batch_size,moe_backend",
[
Expand Down Expand Up @@ -4460,6 +4492,114 @@ def test_eagle3_4gpus(self, moe_backend, one_model, overlap_scheduler,
sampling_params=sampling_params,
extra_evaluator_kwargs=extra_evaluator_kwargs)

@pytest.mark.skip_less_device(4)
@pytest.mark.parametrize("one_model", [True, False],
ids=["one_model", "two_model"])
def test_eagle3_vswa_reuse_4gpus(self, one_model, mocker):
MAX_OUTPUT_LEN = 128179
MAX_INPUT_LEN = 32768

mocker.patch.object(GSM8K, "MAX_OUTPUT_LEN", 8192)
mocker.patch.dict(GSM8K.EVALUATE_KWARGS,
{"scores_filter": "exact_match,flexible-extract"})

mocker.patch.object(GPQADiamond, "MAX_OUTPUT_LEN", MAX_OUTPUT_LEN)
mocker.patch.object(GPQADiamond, "MAX_INPUT_LEN", MAX_INPUT_LEN)

pytorch_config = dict(cuda_graph_config=CudaGraphConfig())
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4,
dtype="auto",
enable_block_reuse=True,
max_attention_window=[128, 32768])

eagle_model_dir = f"{llm_models_root()}/gpt_oss/gpt-oss-120b-Eagle3"
draft_len = 3
spec_config = EagleDecodingConfig(max_draft_len=draft_len,
speculative_model_dir=eagle_model_dir,
eagle3_one_model=one_model,
allow_advanced_sampling=True)

max_seq_len = MAX_INPUT_LEN + MAX_OUTPUT_LEN
llm = LLM(self.MODEL_PATH,
tensor_parallel_size=4,
pipeline_parallel_size=1,
moe_expert_parallel_size=1,
kv_cache_config=kv_cache_config,
max_seq_len=max_seq_len,
speculative_config=spec_config,
**pytorch_config,
enable_attention_dp=False)

with llm:
model_name = "GPT-OSS/120B-MXFP4"

# GSM8K
task = GSM8K(model_name)
task.evaluate(llm,
extra_evaluator_kwargs=self.extra_evaluator_kwargs)

# GPQA Medium Reasoning
task = GPQADiamond(model_name)

chat_template_kwargs = dict(reasoning_effort="medium")
extra_evaluator_kwargs = {
**self.extra_evaluator_kwargs, "chat_template_kwargs":
chat_template_kwargs
}

sampling_params = SamplingParams(
temperature=1.0,
top_p=1.0,
max_tokens=MAX_OUTPUT_LEN,
truncate_prompt_tokens=MAX_INPUT_LEN)

task.evaluate(llm,
sampling_params=sampling_params,
extra_evaluator_kwargs=extra_evaluator_kwargs)

@pytest.mark.skip_less_device(4)
@pytest.mark.parametrize("one_model", [True, False],
ids=["one_model", "two_model"])
def test_eagle3_guided_decoding_4gpus(self, one_model, mocker):
MAX_OUTPUT_LEN = 128179
MAX_INPUT_LEN = 32768

mocker.patch.dict(os.environ, {"TRTLLM_XGUIDANCE_LENIENT": "1"})
mocker.patch.object(GSM8K, "MAX_OUTPUT_LEN", 8192)
mocker.patch.dict(GSM8K.EVALUATE_KWARGS,
{"scores_filter": "exact_match,flexible-extract"})

mocker.patch.object(GPQADiamond, "MAX_OUTPUT_LEN", MAX_OUTPUT_LEN)
mocker.patch.object(GPQADiamond, "MAX_INPUT_LEN", MAX_INPUT_LEN)

pytorch_config = dict(cuda_graph_config=CudaGraphConfig())
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.4,
dtype="auto")

eagle_model_dir = f"{llm_models_root()}/gpt_oss/gpt-oss-120b-Eagle3"
draft_len = 3
spec_config = EagleDecodingConfig(max_draft_len=draft_len,
speculative_model_dir=eagle_model_dir,
eagle3_one_model=one_model,
allow_advanced_sampling=True)

max_seq_len = MAX_INPUT_LEN + MAX_OUTPUT_LEN
llm = LLM(self.MODEL_PATH,
tensor_parallel_size=4,
pipeline_parallel_size=1,
moe_expert_parallel_size=1,
guided_decoding_backend="xgrammar",
kv_cache_config=kv_cache_config,
max_seq_len=max_seq_len,
speculative_config=spec_config,
**pytorch_config,
enable_attention_dp=False)

with llm:
model_name = "GPT-OSS/120B-MXFP4"
task = JsonModeEval(model_name)
task.evaluate(llm)

@pytest.mark.skip_less_device(2)
@pytest.mark.timeout(14400)
@pytest.mark.parametrize("overlap_scheduler", [True, False],
Expand Down
6 changes: 6 additions & 0 deletions tests/integration/test_lists/qa/llm_function_core.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -430,6 +430,8 @@ accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[
accuracy/test_llm_api_pytorch.py::TestLlama3_3_70BInstruct::test_fp8_eagle3_tp8[eagle3_one_model=False-torch_compile=True]
accuracy/test_llm_api_pytorch.py::TestMistral7B::test_auto_dtype
accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_auto_dtype
accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_fp8_vswa_reuse
accuracy/test_llm_api_pytorch.py::TestGemma3_1BInstruct::test_fp8_guided_decoding_vswa_reuse[xgrammar]
accuracy/test_llm_api_pytorch.py::TestGemma3_27BInstruct::test_fp8_prequantized
accuracy/test_llm_api_pytorch.py::TestGemma3_27BInstruct::test_auto_dtype
accuracy/test_llm_api_pytorch.py::TestMistralSmall24B::test_auto_dtype
Expand Down Expand Up @@ -613,6 +615,10 @@ accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-one_model
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-one_model-no_overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-two_model-overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-two_model-no_overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_vswa_reuse_4gpus[one_model]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_vswa_reuse_4gpus[two_model]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_guided_decoding_4gpus[one_model]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_guided_decoding_4gpus[two_model]
accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_auto_dtype[False-False-False]
accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_auto_dtype[True-True-True]
accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ngram
Expand Down
15 changes: 15 additions & 0 deletions tests/integration/test_lists/qa/llm_function_rtx6k.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -151,6 +151,9 @@ accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[ep4-cutlass-auto]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[tp4-cutlass-auto]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4a16[dp4-auto]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_chunked_prefill[cutlass-auto]
accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm]
accuracy/test_llm_api_pytorch.py::TestQwen3_235B_A22B::test_nvfp4[latency_moe_trtllm_eagle3]
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_nvfp4_multi_gpus_sm120[throughput_tp8]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[cutlass-one_model-overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[cutlass-one_model-no_overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[cutlass-two_model-overlap_scheduler]
Expand Down Expand Up @@ -204,6 +207,18 @@ accuracy/test_disaggregated_serving.py::TestLlama4ScoutInstruct::test_auto_dtype
accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_auto_dtype[False]
accuracy/test_disaggregated_serving.py::TestQwen3_8B::test_auto_dtype[True]

accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[trtllm-one_model-overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[trtllm-one_model-no_overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[trtllm-two_model-overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[trtllm-two_model-no_overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-one_model-overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-one_model-no_overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-two_model-overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_4gpus[triton-two_model-no_overlap_scheduler]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_vswa_reuse_4gpus[one_model]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_vswa_reuse_4gpus[two_model]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_guided_decoding_4gpus[one_model]
accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_eagle3_guided_decoding_4gpus[two_model]
test_e2e.py::test_ptp_quickstart_advanced_mixed_precision
test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-BF16-llama-3.1-model/Meta-Llama-3.1-8B]
test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-FP8-llama-3.1-model/Llama-3.1-8B-Instruct-FP8]
Expand Down
Loading
Loading