[Feature][Frontend] Expose effective attention block size for DCP (#56538)

Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
Signed-off-by: Julien Darve <jdarve@NVIDIA.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
JulienDarve
2026-09-16 08:29:43 +00:00
committed by GitHub
co-authored by Bugen Zhao OpenAI Codex
parent 0d8173d153
commit f37c550bf6
18 changed files with 159 additions and 11 deletions
+15 -3
View File
@@ -566,6 +566,7 @@ class MPClient(EngineCoreClient):
):
self.vllm_config = vllm_config
self._renderer: BaseRenderer | None = renderer
self._effective_attention_block_sizes: set[int | None] = set()
# ZMQ setup.
sync_ctx = zmq.Context(io_threads=2)
@@ -803,10 +804,21 @@ class MPClient(EngineCoreClient):
def _apply_ready_response(self, payload: bytes) -> None:
"""Decode an EngineCoreReadyResponse and sync any post-initialization
config changes (e.g. auto-fitted max_model_len) back to the frontend."""
if not payload:
return
vllm_config = self.vllm_config
response = msgspec.msgpack.decode(payload, type=EngineCoreReadyResponse)
response = (
msgspec.msgpack.decode(payload, type=EngineCoreReadyResponse)
if payload
else None
)
self._effective_attention_block_sizes.add(
response.effective_attention_block_size if response is not None else None
)
sizes = self._effective_attention_block_sizes
vllm_config.cache_config.effective_attention_block_size = (
next(iter(sizes)) if len(sizes) == 1 else None
)
if response is None:
return
vllm_config.model_config.max_model_len = min(
vllm_config.model_config.max_model_len, response.max_model_len
)