[PCP][DCP] Support DCP target model with non-DCP Dspark (#56723)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Codex <noreply@openai.com>
This commit is contained in:
Lucas Wilkinson
2026-09-26 18:12:35 -04:00
committed by GitHub
co-authored by Claude Opus 5 Codex
parent 77871126f9
commit 927c87b347
32 changed files with 606 additions and 201 deletions
+34
View File
@@ -1,5 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from types import SimpleNamespace
import pytest
import torch
@@ -7,6 +9,38 @@ from vllm.v1.attention.backends.utils import get_dcp_local_seq_lens
from vllm.v1.worker.cp_utils import should_skip_dcp_context_attention
def test_cp_checks_allow_a_replicated_draft_backend_without_dcp(monkeypatch):
from vllm.v1.worker import cp_utils
config = SimpleNamespace(
parallel_config=SimpleNamespace(
prefill_context_parallel_size=1,
decode_context_parallel_size=4,
cp_kv_cache_interleave_size=16,
),
speculative_config=object(),
)
target = SimpleNamespace(
impl=SimpleNamespace(
dcp_world_size=4,
supports_mtp_with_cp_non_trivial_interleave_size=True,
need_to_return_lse_for_decode=True,
)
)
draft = SimpleNamespace(
impl=SimpleNamespace(
dcp_world_size=1,
supports_mtp_with_cp_non_trivial_interleave_size=False,
need_to_return_lse_for_decode=False,
)
)
layers = {"target": target, "draft": draft}
monkeypatch.setattr(cp_utils, "get_layers_from_vllm_config", lambda *args: layers)
cp_utils.check_attention_cp_compatibility(config, target_layer_names={"target"})
with pytest.raises(AssertionError, match="MTP with cp_kv_cache_interleave_size"):
cp_utils.check_attention_cp_compatibility(config)
def test_skip_gate_only_for_zero_context():
assert should_skip_dcp_context_attention(torch.zeros(3, dtype=torch.int32))
assert not should_skip_dcp_context_attention(