Skip to content

[https://nvbugs/5473781][fix] Fix llama 4 FP8 for PP>1 #7220

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 9 commits into from
Sep 24, 2025
Merged

[https://nvbugs/5473781][fix] Fix llama 4 FP8 for PP>1 #7220

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
dbab58d
[https://nvbugs/5473781][fix] Fix llama 4 FP8 for PP>2
mikeiovine Aug 25, 2025
5a2892f
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 2, 2025
4cfda3d
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 5, 2025
a8d9b15
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 8, 2025
accd194
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 9, 2025
0bf6222
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 10, 2025
84921b3
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 15, 2025
8392a23
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 17, 2025
ea686c3
Merge branch 'main' into fix-llama4-pp
mikeiovine Sep 22, 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
15 changes: 8 additions & 7 deletions tensorrt_llm/_torch/models/modeling_llama.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -419,11 +419,12 @@ def __init__(
overridden_tp_size=1 if self.enable_attention_dp else None,
layer_idx=layer_idx,
)

# TODO(TRTLLM-7809): Fix fusion with PP>1
self.fusion_config.PRE_MLP_FUSION = model_config.mapping.has_tp(
) and not self.enable_attention_dp and self.enable_fusion
self.fusion_config.POST_MLP_FUSION = model_config.mapping.has_tp(
) and not self.enable_attention_dp and self.enable_fusion
) and not self.enable_attention_dp and self.enable_fusion and not model_config.mapping.has_pp(
)
self.fusion_config.POST_MLP_FUSION = self.fusion_config.PRE_MLP_FUSION

else:
self.feed_forward = Llama4MoE(
num_experts=config.num_local_experts,
Expand All @@ -437,9 +438,9 @@ def __init__(
layer_idx=layer_idx)

self.fusion_config.PRE_MOE_FUSION = model_config.mapping.has_tp(
) and not self.enable_attention_dp and self.enable_fusion
self.fusion_config.POST_MOE_FUSION = model_config.mapping.has_tp(
) and not self.enable_attention_dp and self.enable_fusion
) and not self.enable_attention_dp and self.enable_fusion and not model_config.mapping.has_pp(
)
self.fusion_config.POST_MOE_FUSION = self.fusion_config.PRE_MOE_FUSION

self.input_layernorm = RMSNorm(hidden_size=config.hidden_size,
eps=config.rms_norm_eps,
Expand Down
4 changes: 2 additions & 2 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 @@ -696,8 +696,8 @@ def test_chunked_prefill(self, attn_backend):
@parametrize_with_ids("cuda_graph", [False, True])
@pytest.mark.parametrize(
"tp_size,pp_size,ep_size", [(8, 1, 1), (8, 1, 4), (8, 1, 8), (4, 1, 1),
(4, 1, 2), (4, 1, 4)],
ids=["tp8", "tp8ep4", "tp8ep8", "tp4", "tp4ep2", "tp4ep4"])
(4, 1, 2), (4, 1, 4), (4, 2, 1)],
ids=["tp8", "tp8ep4", "tp8ep8", "tp4", "tp4ep2", "tp4ep4", "tp4pp2"])
def test_fp8(self, cuda_graph, tp_size, pp_size, ep_size):
if get_device_memory() < 140000 and get_device_count() < 8:
pytest.skip("Not enough memory for this test")
Expand Down