Skip to content

[None][feat] add qwen3-next CI test of accuracy on BF16 and NVFP4 #9330

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
nv-guomingz merged 5 commits into from
Nov 27, 2025
Merged

[None][feat] add qwen3-next CI test of accuracy on BF16 and NVFP4 #9330

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
f670f1d
add qwen3-next CI test of accuracy on BF16 and NVFP4
JadoTu Nov 20, 2025
dd1d581
add tests in different gpus
JadoTu Nov 21, 2025
c328fe0
remove duplicate tests
JadoTu Nov 23, 2025
dd3d55a
enlarge gsm8k output_token_len and adjust the score
JadoTu Nov 24, 2025
fce137e
add pytest flag
JadoTu Nov 25, 2025
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
5 changes: 5 additions & 0 deletions tests/integration/defs/accuracy/references/gsm8k.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -138,6 +138,11 @@ Qwen3/Qwen3-235B-A22B:
accuracy: 85.78
Qwen3/Qwen3-Next-80B-A3B-Thinking:
- accuracy: 81.577
Qwen3/Qwen3-Next-80B-A3B-Instruct:
- accuracy: 92.72
- quant_algo: NVFP4
kv_cache_quant_algo: FP8
accuracy: 90.86
moonshotai/Kimi-K2-Instruct:
- quant_algo: FP8_BLOCK_SCALES
accuracy: 94.84
Expand Down
5 changes: 5 additions & 0 deletions tests/integration/defs/accuracy/references/mmlu.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -242,6 +242,11 @@ Qwen3/Qwen3-235B-A22B:
accuracy: 86
Qwen3/Qwen3-Next-80B-A3B-Thinking:
- accuracy: 86
Qwen3/Qwen3-Next-80B-A3B-Instruct:
- accuracy: 86.03
- quant_algo: NVFP4
kv_cache_quant_algo: FP8
accuracy: 85.08
moonshotai/Kimi-K2-Instruct:
- quant_algo: FP8_BLOCK_SCALES
accuracy: 87.65
Expand Down
80 changes: 80 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 @@ -4136,6 +4136,86 @@ def test_auto_dtype(self, tp_size, pp_size, ep_size):
task.evaluate(llm)


@skip_pre_hopper
@pytest.mark.skip_less_device_memory(80000)
class TestQwen3NextInstruct(LlmapiAccuracyTestHarness):
MODEL_PATH = f"{llm_models_root()}/Qwen3-Next"
MODEL_NAME = "Qwen3/Qwen3-Next-80B-A3B-Instruct"

# Default setting of `256` is too small
GSM8K_MAX_OUTPUT_LEN = 512

@pytest.mark.skip_less_device(4)
@pytest.mark.parametrize(
"tp_size,pp_size,ep_size,cuda_graph,overlap_scheduler",
[
(4, 1, 4, True, True),
],
ids=[
"tp4ep4_cudagraph_overlap",
],
)
def test_bf16_4gpu(self, tp_size, pp_size, ep_size, cuda_graph,
overlap_scheduler, mocker):
model_path = f"{self.MODEL_PATH}/Qwen3-Next-80B-A3B-Instruct"
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6,
enable_block_reuse=False)
pytorch_config = dict(disable_overlap_scheduler=not overlap_scheduler,
cuda_graph_config=CudaGraphConfig(
max_batch_size=512) if cuda_graph else None)

with LLM(
model_path,
tensor_parallel_size=tp_size,
max_num_tokens=16384,
pipeline_parallel_size=pp_size,
moe_expert_parallel_size=ep_size,
kv_cache_config=kv_cache_config,
**pytorch_config,
) as llm:
task = MMLU(self.MODEL_NAME)
task.evaluate(llm)
mocker.patch.object(GSM8K, "MAX_OUTPUT_LEN",
self.GSM8K_MAX_OUTPUT_LEN)
task = GSM8K(self.MODEL_NAME)
task.evaluate(llm)

@skip_pre_blackwell
@pytest.mark.skip_less_device(4)
@pytest.mark.parametrize("moe_backend", ["CUTLASS", "TRTLLM"],
ids=["cutlass", "trtllm"])
@pytest.mark.parametrize(
"tp_size,pp_size,ep_size,cuda_graph,overlap_scheduler",
[(1, 1, 1, True, True), (4, 1, 1, True, True), (4, 1, 4, True, True),
(4, 1, 4, False, False)],
ids=["tp1", "tp4ep1", "tp4ep4", "no_cuda_graph_overlap"])
def test_nvfp4(self, moe_backend, tp_size, pp_size, ep_size, cuda_graph,
overlap_scheduler, mocker):
model_path = f"{self.MODEL_PATH}/qwen3-next-80b-instruct-nvfp4-ptq-fp8kv"

kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.6,
enable_block_reuse=False)
pytorch_config = dict(disable_overlap_scheduler=not overlap_scheduler,
cuda_graph_config=CudaGraphConfig(
max_batch_size=512) if cuda_graph else None)
moe_config = MoeConfig(backend=moe_backend)

with LLM(model_path,
tensor_parallel_size=tp_size,
max_num_tokens=16384,
pipeline_parallel_size=pp_size,
moe_expert_parallel_size=ep_size,
kv_cache_config=kv_cache_config,
**pytorch_config,
moe_config=moe_config) as llm:
task = MMLU(self.MODEL_NAME)
task.evaluate(llm)
mocker.patch.object(GSM8K, "MAX_OUTPUT_LEN",
self.GSM8K_MAX_OUTPUT_LEN)
task = GSM8K(self.MODEL_NAME)
task.evaluate(llm)


class TestSeedOss_36B(LlmapiAccuracyTestHarness):
MODEL_NAME = "ByteDance-Seed/Seed-OSS-36B-Instruct"
MODEL_PATH = f"{llm_models_root()}/Seed-OSS/Seed-OSS-36B-Instruct"
Expand Down
5 changes: 5 additions & 0 deletions tests/integration/test_lists/test-db/l0_b200.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,6 +54,11 @@ l0_b200:
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[mxfp8-latency-TRTLLM]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a8_mxfp4[mxfp8-latency-CUTLASS]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_w4a16_mxfp4[latency-TRTLLM]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[no_cuda_graph_overlap-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-trtllm]
- disaggregated/test_workers.py::test_workers_kv_cache_aware_router_eviction[TinyLlama-1.1B-Chat-v1.0] # nvbugs 5300551
- test_e2e.py::test_ptp_quickstart_advanced[Llama3.1-8B-NVFP4-nvfp4-quantized/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
6 changes: 6 additions & 0 deletions tests/integration/test_lists/test-db/l0_dgx_b200.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -42,6 +42,12 @@ l0_dgx_b200:
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=False]
- accuracy/test_llm_api_pytorch.py::TestQwen3_30B_A3B::test_nvfp4[dep4_latency_moe_cutlass-torch_compile=True] ISOLATION
- accuracy/test_llm_api_pytorch.py::TestQwen3NextThinking::test_auto_dtype[tp4ep4]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_bf16_4gpu[tp4ep4_cudagraph_overlap]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[no_cuda_graph_overlap-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-trtllm]
- 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]
Expand Down
5 changes: 5 additions & 0 deletions tests/integration/test_lists/test-db/l0_dgx_b300.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -58,6 +58,11 @@ l0_dgx_b300:
- 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=True]
- accuracy/test_disaggregated_serving.py::TestQwen3_30B_A3B::test_mixed_ctx_gen_model[ctxpp2gentp2]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[no_cuda_graph_overlap-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-trtllm]
- 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-cutlass-fp8]
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[tp4-cutlass-auto]
Expand Down
1 change: 1 addition & 0 deletions tests/integration/test_lists/test-db/l0_dgx_h100.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -93,6 +93,7 @@ l0_dgx_h100:
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ctx_pp_gen_tp_asymmetric[MMLU-gen_tp=2-ctx_pp=2]
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_multi_instance[GSM8K]
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_multi_instance[MMLU]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_bf16_4gpu[tp4ep4_cudagraph_overlap]
- disaggregated/test_auto_scaling.py::test_service_discovery[etcd-round_robin]
- disaggregated/test_auto_scaling.py::test_service_discovery[etcd-load_balancing]
- disaggregated/test_auto_scaling.py::test_worker_restart[etcd-round_robin]
Expand Down
1 change: 1 addition & 0 deletions tests/integration/test_lists/test-db/l0_dgx_h200.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -33,6 +33,7 @@ l0_dgx_h200:
- accuracy/test_disaggregated_serving.py::TestLlama3_1_8BInstruct::test_ctx_pp_gen_tp_asymmetric[MMLU-gen_tp=2-ctx_pp=4]
- accuracy/test_disaggregated_serving.py::TestGPTOSS::test_auto_dtype[False]
- accuracy/test_disaggregated_serving.py::TestGPTOSS::test_auto_dtype[True]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_bf16_4gpu[tp4ep4_cudagraph_overlap]
- disaggregated/test_disaggregated.py::test_disaggregated_ctxtp2pp2_gentp2pp2[TinyLlama-1.1B-Chat-v1.0]
- disaggregated/test_disaggregated.py::test_disaggregated_ctxpp4_genpp4[TinyLlama-1.1B-Chat-v1.0]
- unittest/llmapi/test_llm_pytorch.py::test_nemotron_nas_lora
Expand Down
5 changes: 5 additions & 0 deletions tests/integration/test_lists/test-db/l0_gb200_multi_gpus.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,11 @@ l0_gb200_multi_gpus:
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[tp4-trtllm-auto]
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus[dp4-cutlass-auto]
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_4gpus_online_eplb[fp8]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[no_cuda_graph_overlap-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-trtllm]
- condition:
ranges:
system_gpu_count:
Expand Down
2 changes: 2 additions & 0 deletions tests/integration/test_lists/test-db/l0_rtx_pro_6000.yml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -43,6 +43,8 @@ l0_rtx_pro_6000:
- accuracy/test_llm_api_pytorch.py::TestGPTOSS::test_w4_1gpu[True-True-cutlass-auto]
- accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_fp4
- accuracy/test_llm_api_pytorch.py::TestPhi4MM::test_fp8
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp1-cutlass]
- accuracy/test_llm_api_pytorch.py::TestQwen3NextInstruct::test_nvfp4[tp4ep4-trtllm]

- condition:
ranges:
Expand Down