[PCP][DCP] Enable PCP+DCP on sparse-MLA models (#56157)

Signed-off-by: patryk.saffer <patryk.saffer@blc-login-1.blc1.mistralcompute.ai>
Signed-off-by: patryk.saffer <patryk.saffer@blc-login-0.blc1.mistralcompute.ai>
Signed-off-by: PatrykSaffer <patryk.saffer@mistral.ai>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: patryk.saffer <patryk.saffer@blc-login-1.blc1.mistralcompute.ai>
Co-authored-by: patryk.saffer <patryk.saffer@blc-login-0.blc1.mistralcompute.ai>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Patryk Saffer <PatrykSaffer@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
PatrykSaffer
2026-09-12 23:24:44 +00:00
committed by GitHub
co-authored by patryk.saffer patryk.saffer Lucas Wilkinson Patryk Saffer Claude Fable 5.1
parent 72d4d83157
commit ebe1dec2da
21 changed files with 917 additions and 138 deletions
@@ -1080,6 +1080,7 @@ def test_flashmla_forward_bf16_kv_slices_req_id_to_mqa_tokens():
_convert_logical_to_physical_topk=_convert_topk,
index_group=None,
index_group_index=0,
pcp_dcp_kv_gather=False,
)
out, _ = FlashMLASparseImpl._forward_bf16_kv(
@@ -3257,6 +3258,7 @@ def test_flashmla_fp8_metadata_excludes_zero_token_decode_padding(monkeypatch):
device=torch.device(DEVICE_TYPE),
dummy_block_table=torch.zeros(7, 1, device=DEVICE_TYPE),
max_model_len_tensor=torch.zeros(7, device=DEVICE_TYPE),
pcp_dcp_kv_gather=False,
)
query_start_loc_cpu = torch.tensor([0, 110, 220, 330, 440, 550, 660, 660])
common_metadata = SimpleNamespace(
@@ -3363,6 +3365,7 @@ def test_flashmla_fp8_paths_accept_decode_subset(monkeypatch, use_mixed_batch: b
index_group=None,
index_group_index=0,
dcp_world_size=1,
pcp_dcp_kv_gather=False,
need_to_return_lse_for_decode=False,
_fp8_flash_mla_kernel=run_kernel,
_convert_logical_to_physical_topk=(