Skip to content

[https://nvbugs/6412108][fix] AutoDeploy:Shard shared Qwen3.5 experts #16286

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
greg-kwasniewski1 merged 1 commit into from
Jul 16, 2026
Merged

[https://nvbugs/6412108][fix] AutoDeploy:Shard shared Qwen3.5 experts #16286

Show file tree
Hide file tree
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
2 changes: 0 additions & 2 deletions examples/auto_deploy/model_registry/configs/qwen3.5_moe_400b.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,8 +36,6 @@ transforms:
apply_sharding_hints:
enabled: true
allreduce_strategy: SYMM_MEM
# Shared expert is excluded from sharding for performance purpose
shard_layers: ["moe", "delta", "mha"]
simple_shard_filter: "lm_head"
multi_stream_moe:
stage: compile
Expand Down
10 changes: 8 additions & 2 deletions tensorrt_llm/_torch/auto_deploy/models/custom/modeling_qwen3_5_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -635,25 +635,31 @@ def __init__(self, config: Qwen3_5MoeTextConfig, intermediate_size: int):
self.act_fn = ACT2FN[config.hidden_act]

def forward(self, x: torch.Tensor) -> torch.Tensor:
# Intentionally left untagged: with no ``layer_type`` it defaults to "unknown",
# which the ``shard_layers`` inclusion whitelist excludes -> kept replicated.
# Tagged ``layer_type="mlp"`` so the shared expert is TP-sharded like the
# routed experts (colwise gate/up + rowwise down). Its rowwise down_proj
# therefore emits a per-rank partial that is summed with the routed
# partial and lifted to full by the single merge-point all_reduce in
# ``Qwen3_5MoeSparseMoeBlock.forward``.
gate = torch.ops.auto_deploy.torch_linear_simple(
x,
self.gate_proj.weight,
self.gate_proj.bias,
tp_mode="colwise",
layer_type="mlp",
)
up = torch.ops.auto_deploy.torch_linear_simple(
x,
self.up_proj.weight,
self.up_proj.bias,
tp_mode="colwise",
layer_type="mlp",
)
return torch.ops.auto_deploy.torch_linear_simple(
self.act_fn(gate) * up,
self.down_proj.weight,
self.down_proj.bias,
tp_mode="rowwise",
layer_type="mlp",
)


Expand Down
1 change: 0 additions & 1 deletion tests/integration/test_lists/waives.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,6 @@ accuracy/test_epd_disagg_multimodal.py::TestVideoMMEEPD::test_disaggregated_vide
accuracy/test_llm_api.py::TestLlama3_1_8BInstruct::test_guided_decoding_4gpus[xgrammar] SKIP (https://nvbugs/5346443)
accuracy/test_llm_api_autodeploy.py::TestMiniMaxM2::test_finegrained_fp8 SKIP (https://nvbugs/6396422)
accuracy/test_llm_api_autodeploy.py::TestNemotronSuperV3::test_mtp[nvfp4_ws8_80gb-trtllm] SKIP (https://nvbugs/6450341)
accuracy/test_llm_api_autodeploy.py::TestQwen3_5_397B_MoE::test_nvfp4[8] SKIP (https://nvbugs/6412108)
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[throughput_mtp] SKIP (https://nvbugs/6428101)
accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[throughput_mtp_trtllm] SKIP (https://nvbugs/6426868)
accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_dsa_host_cache_offload[host_cache_offload] SKIP (https://nvbugs/6384136)
Expand Down
Loading