apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[BugFix] fix multi-step mtp bug (#6754)

Browse Source
This commit is contained in:
ddchenhao66
2026-03-11 10:16:04 +08:00
committed by GitHub
parent b05a6c4206
commit a502dda1fe
2 changed files with 4 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/engine/sched/resource_manager_v1.py
+4 -5
View File
@@ -229,6 +229,8 @@ class ResourceManagerV1(ResourceManager):
if self.config.speculative_config.method is not None:
block_num = min(block_num + 1, self.config.cache_config.max_block_num_per_seq)
else:
block_num = min(block_num, self.config.cache_config.max_block_num_per_seq)
return block_num
def _prepare_prefill_task(self, request, new_token_num):
@@ -926,7 +928,7 @@ class ResourceManagerV1(ResourceManager):
)
# Allocate blocks to prefill
if self.cache_manager.can_allocate_gpu_blocks(can_schedule_block_num_threshold):
if not request.get("skip_allocate", False):
if num_new_block > 0:
extra_gpu_block_ids = self.cache_manager.allocate_gpu_blocks(
num_new_block, request.request_id
)
@@ -985,7 +987,7 @@ class ResourceManagerV1(ResourceManager):
)
# Allocate blocks to prefill
if self.cache_manager.can_allocate_gpu_blocks(can_schedule_block_num_threshold):
if not request.get("skip_allocate", False):
if num_new_block > 0:
extra_gpu_block_ids = self.cache_manager.allocate_gpu_blocks(
num_new_block, request.request_id
)
@@ -1166,19 +1168,16 @@ class ResourceManagerV1(ResourceManager):
request.cache_info = [matched_block_num, no_cache_block_num]
request.block_tables = common_block_ids
request.skip_allocate = False
request.num_cached_tokens = matched_token_num
if self.config.cache_config.disable_chunked_mm_input:
if matched_token_num == request.need_prefill_tokens:
matched_token_num = matched_token_num - self.config.cache_config.block_size
request.skip_allocate = True
request.num_computed_tokens = self.revert_chunked_mm_input(
request.multimodal_inputs, matched_token_num
)
else:
if matched_token_num == request.need_prefill_tokens:
request.num_computed_tokens = matched_token_num - self.config.cache_config.block_size
request.skip_allocate = True
else:
request.num_computed_tokens = matched_token_num