[SimpleCPUOffloadConnector] Fix remaining global→block conversions under PCP/DCP (#46394)

Signed-off-by: Jeff Ma <jeffjma@umich.edu>
This commit is contained in:
Jeff (Junze) Ma
2026-06-24 23:05:24 -07:00
committed by GitHub
parent fc61c6fc26
commit 36fd7e8b86
2 changed files with 94 additions and 6 deletions
+13 -6
View File
@@ -103,9 +103,12 @@ class SimpleCPUOffloadScheduler:
assert 0 <= self.fa_gidx < len(self.cpu_kv_cache_config.kv_cache_groups)
# FA group's own block_size; divides scheduler_block_size (the LCM)
# but is NOT assumed to equal it.
self.fa_block_size: int = self.cpu_kv_cache_config.kv_cache_groups[
self.fa_gidx
].kv_cache_spec.block_size
self.fa_block_size: int = (
self.cpu_kv_cache_config.kv_cache_groups[
self.fa_gidx
].kv_cache_spec.block_size
* self.cp_world_size
)
assert self.block_size % self.fa_block_size == 0
logger.info(
@@ -348,10 +351,12 @@ class SimpleCPUOffloadScheduler:
# the rest will be released along with the temp pin below.
cpu_hit_blocks: list[list[KVCacheBlock]] = []
for g in range(num_groups):
g_block_size = kv_cache_groups[g].kv_cache_spec.block_size
g_block_size = (
kv_cache_groups[g].kv_cache_spec.block_size * self.cp_world_size
)
assert num_external_tokens % g_block_size == 0, (
f"num_external_tokens={num_external_tokens} not aligned to "
f"group {g} block_size={g_block_size}"
f"group {g} effective block_size={g_block_size}"
)
n_take_g = num_external_tokens // g_block_size
cpu_hit_blocks.append(cpu_hit_blocks_full[g][:n_take_g])
@@ -599,7 +604,9 @@ class SimpleCPUOffloadScheduler:
already_stored_g = state.num_stored_blocks[g]
group_gpu_ids = block_ids_by_group[g]
g_block_size = kv_cache_groups[g].kv_cache_spec.block_size
g_block_size = (
kv_cache_groups[g].kv_cache_spec.block_size * self.cp_world_size
)
ready_blocks_g = aligned_tokens // g_block_size
scannable = group_gpu_ids[already_stored_g:ready_blocks_g]