[Model] Support DeepSeek-V4.1-Flash (#56214)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com> Signed-off-by: Jee Jee Li <jeejeelee@inferact.ai> Signed-off-by: andyluo7 <andy.luo@amd.com> Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> Co-authored-by: Jee Jee Li <jeejeelee@inferact.ai> Co-authored-by: andyluo7 <andy.luo@amd.com> Co-authored-by: OpenAI Codex <codex@openai.com> Co-authored-by: zjy0516 <riverclouds.zhu@qq.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Jee Jee Li
andyluo7
OpenAI Codex
zjy0516
Claude Opus 5
parent
e6821ceac9
commit
e77daef89e
@@ -51,7 +51,7 @@ from vllm.v1.simple_kv_offload.metadata import (
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.v1.core.kv_cache_manager import KVCacheBlocks
|
||||
from vllm.v1.core.kv_cache_utils import KVCacheBlock
|
||||
from vllm.v1.core.kv_cache_utils import BlockHashList, KVCacheBlock
|
||||
from vllm.v1.kv_cache_interface import KVCacheConfig
|
||||
from vllm.v1.request import Request
|
||||
|
||||
@@ -783,6 +783,18 @@ class SimpleCPUOffloadScheduler:
|
||||
req_ids = list(dict.fromkeys(req_ids))
|
||||
return merged_gpu_block_ids, merged_cpu_block_ids, req_ids, merged_block_meta
|
||||
|
||||
def _cached_gpu_block(
|
||||
self, resolved_hashes: "BlockHashList", block_idx: int, group_id: int
|
||||
) -> "KVCacheBlock | None":
|
||||
"""Return the GPU block still cached under this group's block hash."""
|
||||
if block_idx >= len(resolved_hashes):
|
||||
return None
|
||||
assert self._gpu_block_pool is not None
|
||||
blocks = self._gpu_block_pool.get_cached_block(
|
||||
resolved_hashes[block_idx], [group_id]
|
||||
)
|
||||
return blocks[0] if blocks else None
|
||||
|
||||
def _select_eager_blocks_to_store(
|
||||
self,
|
||||
state: StoreRequestState,
|
||||
@@ -827,18 +839,25 @@ class SimpleCPUOffloadScheduler:
|
||||
# the middle of the request.
|
||||
group_size = self.group_block_sizes[g]
|
||||
ready = min(len(group_gpu_ids), aligned_tokens // group_size)
|
||||
resolved_hashes = (
|
||||
resolve_block_hashes(
|
||||
request.block_hashes, self.hash_block_size, group_size
|
||||
)
|
||||
if block_meta is not None
|
||||
else None
|
||||
resolved_hashes = resolve_block_hashes(
|
||||
request.block_hashes, self.hash_block_size, group_size
|
||||
)
|
||||
curr_mm_idx = 0
|
||||
secondary_mm_idx = 0
|
||||
start = state.num_stored_blocks[g]
|
||||
for i, gpu_block_id in enumerate(group_gpu_ids[start:ready], start=start):
|
||||
gpu_block = self._gpu_block_pool.blocks[gpu_block_id]
|
||||
if gpu_block.is_null:
|
||||
# Sliding-window groups null pages that left the window
|
||||
# before the connector sees the block table, but the
|
||||
# retained prefix-cache tail stays hashed in the GPU free
|
||||
# queue; store it from there.
|
||||
recovered = self._cached_gpu_block(resolved_hashes, i, g)
|
||||
if recovered is None:
|
||||
advanced_per_group[g] += 1
|
||||
continue
|
||||
gpu_block = recovered
|
||||
gpu_block_id = gpu_block.block_id
|
||||
if (
|
||||
self._classify_store_candidate(gpu_block)
|
||||
is not _StoreAdmission.READY
|
||||
@@ -854,7 +873,6 @@ class SimpleCPUOffloadScheduler:
|
||||
if block_meta is not None:
|
||||
token_start = i * group_size
|
||||
token_end = token_start + group_size
|
||||
assert resolved_hashes is not None
|
||||
parent_hash = (
|
||||
None
|
||||
if i == 0
|
||||
|
||||
Reference in New Issue
Block a user