Skip to content

fix: [nvbugs/5351130] Adjust DSV3-Lite tests free_gpu_memory_fraction to 0.75 to prevent OOM on CI. #5896

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
merged 1 commit into from
Jul 10, 2025
Merged

fix: [nvbugs/5351130] Adjust DSV3-Lite tests free_gpu_memory_fraction to 0.75 to prevent OOM on CI. #5896

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
20 changes: 10 additions & 10 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 @@ -510,7 +510,7 @@ def test_bfloat16(self, mtp_nextn, attention_dp, cuda_graph,
pytest.skip("https://nvbugs/5252313")
if torch_compile and attention_dp:
pytest.skip("https://nvbugs/5252559")
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
torch_compile_config = TorchCompileConfig(
enable_fullgraph=True,
enable_piecewise_cuda_graph=cuda_graph) if torch_compile else None
Expand Down Expand Up @@ -556,7 +556,7 @@ def test_bfloat16_4gpus(self, tp_size, pp_size, ep_size, mtp_nextn,
pytest.skip("https://nvbugs/5252559")
if torch_compile and pp_size > 1:
pytest.skip("PP with torch.compile is not supported yet.")
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
torch_compile_config = TorchCompileConfig(
enable_fullgraph=True,
enable_piecewise_cuda_graph=cuda_graph) if torch_compile else None
Expand Down Expand Up @@ -600,7 +600,7 @@ def test_fp8_block_scales(self, mtp_nextn, fp8kv, attention_dp, cuda_graph,
pytest.skip("https://nvbugs/5252313")
if torch_compile and attention_dp:
pytest.skip("https://nvbugs/5252559")
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
torch_compile_config = TorchCompileConfig(
enable_fullgraph=True,
enable_piecewise_cuda_graph=cuda_graph) if torch_compile else None
Expand Down Expand Up @@ -642,7 +642,7 @@ def test_fp8_block_scales(self, mtp_nextn, fp8kv, attention_dp, cuda_graph,
@pytest.mark.skip_device_not_contain(["H100"])
@parametrize_with_ids("mtp_nextn", [0, 2])
def test_fp8_block_scales_cuda_graph_padding(self, mtp_nextn):
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
mtp_config = None
if mtp_nextn > 0:
mtp_config = MTPDecodingConfig(num_nextn_predict_layers=mtp_nextn)
Expand All @@ -669,7 +669,7 @@ def test_fp8_block_scales_cuda_graph_padding(self, mtp_nextn):
@parametrize_with_ids("attention_dp", [False, True])
def test_fp8_block_scales_cuda_graph_padding_4gpus(self, mtp_nextn,
attention_dp):
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
mtp_config = None
if mtp_nextn > 0:
mtp_config = MTPDecodingConfig(num_nextn_predict_layers=mtp_nextn)
Expand Down Expand Up @@ -721,7 +721,7 @@ def test_fp8_block_scales_4gpus(self, tp_size, pp_size, ep_size, mtp_nextn,
pytest.skip("https://nvbugs/5252559")
if torch_compile and pp_size > 1:
pytest.skip("PP with torch.compile is not supported yet.")
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
torch_compile_config = TorchCompileConfig(
enable_fullgraph=True,
enable_piecewise_cuda_graph=cuda_graph) if torch_compile else None
Expand Down Expand Up @@ -766,7 +766,7 @@ def test_fp8_block_scales_4gpus(self, tp_size, pp_size, ep_size, mtp_nextn,
@pytest.mark.skip_less_device(4)
@pytest.mark.skip_device_not_contain(["H100", "H200"])
def test_fp8_block_scales_4gpus_static_eplb(self):
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)

num_experts = 72
num_slots = 80
Expand Down Expand Up @@ -813,7 +813,7 @@ def test_nvfp4(self, fp8kv, attention_dp, cuda_graph, overlap_scheduler,
pytest.skip("https://nvbugs/5252313")
if torch_compile and attention_dp:
pytest.skip("https://nvbugs/5252559")
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
torch_compile_config = TorchCompileConfig(
enable_fullgraph=True,
enable_piecewise_cuda_graph=cuda_graph) if torch_compile else None
Expand Down Expand Up @@ -879,7 +879,7 @@ def test_nvfp4_4gpus(self, fp8kv, attention_dp, cuda_graph,
if torch_compile and pp_size > 1:
pytest.skip("PP with torch.compile is not supported yet.")

kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9)
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75)
torch_compile_config = TorchCompileConfig(
enable_fullgraph=True,
enable_piecewise_cuda_graph=cuda_graph) if torch_compile else None
Expand Down Expand Up @@ -946,7 +946,7 @@ def test_no_kv_cache_reuse(self, quant_dtype, mtp_nextn, fp8kv,
elif quant_dtype == "nvfp4":
model_path = f"{llm_models_root()}/DeepSeek-V3-Lite/nvfp4_moe_only"

kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.9,
kv_cache_config = KvCacheConfig(free_gpu_memory_fraction=0.75,
enable_block_reuse=False)
pytorch_config = dict(
disable_overlap_scheduler=not overlap_scheduler,
Expand Down