Skip to content

[None][fix] fix TP support for DeepSeek-V3.2 on hopper #9484

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
lfr-0531 merged 3 commits into from
Nov 27, 2025
Merged

[None][fix] fix TP support for DeepSeek-V3.2 on hopper #9484

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0966676
fix TP support for DeepSeek-V3.2 on hopper.
lfr-0531 Nov 26, 2025
0971624
Merge branch 'main' into user/fanrongl/fix_ds32_tp_sm90
lfr-0531 Nov 27, 2025
337fc31
update dsv32 test.
lfr-0531 Nov 27, 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
3 changes: 2 additions & 1 deletion tensorrt_llm/_torch/modules/attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2060,9 +2060,10 @@ def forward_sparse_mla_kvcache_bf16(

# [seq, num_heads, kv_lora_rank], account for padding
attn_out_latent = attn_out_latent[:, :self.num_heads_tp, :]
# TODO: seems we need .contiguous() here when padding enabled before pass to bmm?
attn_out_latent = attn_out_latent.view(
[-1, self.num_heads_tp, self.kv_lora_rank])
if self.num_heads_tp != padding:
attn_out_latent = attn_out_latent.contiguous()

assert (attn_out_latent.shape[0] == q.shape[0]
and attn_out_latent.shape[1] == self.num_heads_tp)
Expand Down
6 changes: 5 additions & 1 deletion 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 @@ -2409,8 +2409,12 @@ class TestDeepSeekV32(LlmapiAccuracyTestHarness):
(8, 1, 8, 1, False, True, True, True, 24, "_DEFAULT"),
(8, 1, 8, 0, True, True, True, True, 24, "_DEFAULT"),
(8, 1, 8, 3, False, False, True, True, 1, "TRTLLM"),
(8, 1, 8, 3, False, False, True, True, 1, "_DEFAULT"),
],
ids=["baseline", "baseline_mtp1", "baseline_fp8kv", "latency"])
ids=[
"baseline", "baseline_mtp1", "baseline_fp8kv", "latency",
"latency_default"
])
def test_fp8_blockscale(self, tp_size, pp_size, ep_size, mtp_nextn, fp8kv,
attention_dp, cuda_graph, overlap_scheduler,
max_batch_size, moe_backend):
Expand Down
2 changes: 1 addition & 1 deletion 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 @@ -18,7 +18,7 @@ l0_dgx_h200:
# - accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[throughput] # OOM
- accuracy/test_llm_api_pytorch.py::TestDeepSeekR1::test_fp8_blockscale[latency] # 1h
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_fp8_blockscale[baseline]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_fp8_blockscale[latency]
- accuracy/test_llm_api_pytorch.py::TestDeepSeekV32::test_fp8_blockscale[latency_default]
- accuracy/test_disaggregated_serving.py::TestLlama4ScoutInstruct::test_auto_dtype[True]
- accuracy/test_disaggregated_serving.py::TestLlama4ScoutInstruct::test_auto_dtype[False]
- accuracy/test_disaggregated_serving.py::TestDeepSeekV3Lite::test_auto_dtype[mtp_nextn=0-overlap_scheduler=True]
Expand Down