[1/N][Cleanup] Standardize on use of is_quantized_kv_cache (#38659)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
@@ -16,6 +16,7 @@ from vllm.model_executor.layers.attention.mla_attention import (
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.platforms.interface import DeviceCapability
|
||||
from vllm.utils.platform_utils import num_compute_units
|
||||
from vllm.utils.torch_utils import is_quantized_kv_cache
|
||||
from vllm.v1.attention.backend import (
|
||||
AttentionBackend,
|
||||
AttentionCGSupport,
|
||||
@@ -571,7 +572,7 @@ class FlashMLASparseImpl(SparseMLAAttentionImpl[FlashMLASparseMetadata]):
|
||||
vllm_config = get_current_vllm_config()
|
||||
max_tokens = vllm_config.scheduler_config.max_num_batched_tokens
|
||||
q_concat_shape = (max_tokens, num_heads, head_size)
|
||||
if kv_cache_dtype.startswith("fp8"):
|
||||
if is_quantized_kv_cache(kv_cache_dtype):
|
||||
assert kv_cache_dtype == "fp8_ds_mla", (
|
||||
"FlashMLA Sparse Attention backend fp8 only supports "
|
||||
"fp8_ds_mla kv-cache dtype"
|
||||
|
||||
Reference in New Issue
Block a user