[Bugfix]Fix DeepSeek-V4 fp8_ds_mla KV cache reshape (#47716)

Co-authored-by: yy-fighting <23518844576@qq.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
Zhiyi Hong
2026-07-06 12:56:44 -07:00
committed by GitHub
co-authored by yy-fighting Michael Goin
parent ae098abe3f
commit 04adc8843b
3 changed files with 14 additions and 2 deletions
+6 -1
View File
@@ -7164,7 +7164,12 @@ class GPUModelRunner(
layer_cache_dtype_str = (
"auto"
if kv_cache_spec.kv_quant_mode == KVQuantMode.NONE
else self.cache_config.cache_dtype
else getattr(
kv_cache_spec,
"cache_dtype_str",
None,
)
or self.cache_config.cache_dtype
)
kv_cache_shape = attn_backend.get_kv_cache_shape(
kernel_num_blocks,