[Hisparse] fix: account for MiniMax HiSparse full-pool memory (#40743)

This commit is contained in:
Shuwen Wang
2026-09-23 13:13:48 +08:00
committed by GitHub
parent 9d267655e3
commit ced0a2c98c
@@ -389,8 +389,18 @@ class DefaultPoolConfigurator(MemoryPoolConfigurator):
indexer_head_dim = sparse_cfg["sparse_index_dim"]
indexer_dtype_size = torch._utils._element_size(kvc.model_dtype)
full_pool_ratio = 1
if get_memory().enable_hisparse:
from sglang.srt.mem_cache.sparsity import parse_hisparse_config
full_pool_ratio = parse_hisparse_config().host_to_device_ratio
main_pool_bytes = (
(num_dense + num_sparse) * 2 * kv_heads * head_dim * kv_size
(num_dense * full_pool_ratio + num_sparse)
* 2
* kv_heads
* head_dim
* kv_size
)
indexer_bytes = (
(num_indexer_kv * 2 + num_indexer_k_only)
@@ -399,7 +409,7 @@ class DefaultPoolConfigurator(MemoryPoolConfigurator):
)
# FP4 scale buffer adjustment doesn't apply to MiniMax sparse:
# cell_size is already a sum over heterogeneous sub-pools.
return main_pool_bytes + indexer_bytes
return main_pool_bytes + indexer_bytes * full_pool_ratio
else:
n = model_config.get_num_kv_heads(tp_size, dcp_size)
cell_size = (