[Config] Unify indexer cache dtype under attention_config.indexer_kv_dtype (#52550)
Signed-off-by: Yongye Zhu <zyy1102000@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
0ad04cff1b
commit
967e104fad
@@ -498,7 +498,9 @@ class MiniMaxM3SparseAttention(nn.Module, AttentionLayerBase):
|
||||
set_default_quant_scales(self, register_buffer=True)
|
||||
# Indexer side-cache dtype, mirroring --kv-cache-dtype for the main
|
||||
# cache (--attention-config '{"indexer_kv_dtype": ...}').
|
||||
self.indexer_kv_dtype = vllm_config.attention_config.indexer_kv_dtype
|
||||
self.indexer_kv_dtype = vllm_config.attention_config.resolve_indexer_kv_dtype(
|
||||
"bf16"
|
||||
)
|
||||
|
||||
# Shared top-k buffer: the indexer writes the selected blocks into it and
|
||||
# the attend impl reads them back (so nothing crosses the eager break as a
|
||||
|
||||
Reference in New Issue
Block a user