[PD] Fix clip logic when state indices lens are mismatch (#23323)
This commit is contained in:
@@ -1003,16 +1003,20 @@ class MooncakeKVManager(CommonKVManager):
|
|||||||
raise RuntimeError(
|
raise RuntimeError(
|
||||||
f"PD Disaggregation does NOT support PD different TP sizes for non-MLA {state_type.upper()} hybrid models yet."
|
f"PD Disaggregation does NOT support PD different TP sizes for non-MLA {state_type.upper()} hybrid models yet."
|
||||||
)
|
)
|
||||||
if len(prefill_state_indices) < len(req.dst_state_indices):
|
dst_state_indices = req.dst_state_indices
|
||||||
|
if len(prefill_state_indices) > len(dst_state_indices):
|
||||||
logger.warning(
|
logger.warning(
|
||||||
f"len(prefill_state_indices) = {len(prefill_state_indices)}, len(dst_state_indices) = {len(req.dst_state_indices)}"
|
f"len(prefill_state_indices) = {len(prefill_state_indices)}, len(dst_state_indices) = {len(dst_state_indices)}"
|
||||||
)
|
)
|
||||||
prefill_state_indices = prefill_state_indices[
|
prefill_state_indices = prefill_state_indices[: len(dst_state_indices)]
|
||||||
: len(req.dst_state_indices)
|
elif len(prefill_state_indices) < len(dst_state_indices):
|
||||||
]
|
logger.warning(
|
||||||
|
f"len(prefill_state_indices) = {len(prefill_state_indices)}, len(dst_state_indices) = {len(dst_state_indices)}"
|
||||||
|
)
|
||||||
|
dst_state_indices = dst_state_indices[: len(prefill_state_indices)]
|
||||||
# Reuse _send_kvcache_generic interface to send extra pool data
|
# Reuse _send_kvcache_generic interface to send extra pool data
|
||||||
prefill_state_indices = np.array(prefill_state_indices, dtype=np.int32)
|
prefill_state_indices = np.array(prefill_state_indices, dtype=np.int32)
|
||||||
dst_state_indices = np.array(req.dst_state_indices, dtype=np.int32)
|
dst_state_indices = np.array(dst_state_indices, dtype=np.int32)
|
||||||
return self._send_kvcache_generic(
|
return self._send_kvcache_generic(
|
||||||
mooncake_session_id=req.mooncake_session_id,
|
mooncake_session_id=req.mooncake_session_id,
|
||||||
src_data_ptrs=self.kv_args.state_data_ptrs,
|
src_data_ptrs=self.kv_args.state_data_ptrs,
|
||||||
|
|||||||
Reference in New Issue
Block a user