A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 22:35 UTC
[UX] Rename CUTLASS_MLA_VLLM_V1 to CUTLASS_MLA (#21966)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Jul 30, 2025 at 23:40 UTC
61445453df8e514d9ddf4d6bd3f9063f120cdac5
4 files changed
+8
-8
vllm/engine/arg_utils.py
+1
-1
@@ -1417,7 +1417,7 @@ class EngineArgs:
1417
"PALLAS_VLLM_V1",
1418
"TRITON_ATTN_VLLM_V1",
1419
"TRITON_MLA",
1420
- "CUTLASS_MLA_VLLM_V1",
1420
+ "CUTLASS_MLA",
1421
"FLASHMLA",
1422
"FLASHINFER",
1423
"FLASHINFER_VLLM_V1",
vllm/platforms/cuda.py
+5
-5
@@ -162,7 +162,7 @@ class CudaPlatformBase(Platform):
162
if cls.is_device_capability(100):
163
# Blackwell => Force CutlassMLA.
164
use_cutlass_mla = True
165
- envs.VLLM_ATTENTION_BACKEND = "CUTLASS_MLA_VLLM_V1"
165
+ envs.VLLM_ATTENTION_BACKEND = "CUTLASS_MLA"
166
else:
167
# Not Blackwell
168
use_flashmla = True
@@ -170,7 +170,7 @@ class CudaPlatformBase(Platform):
170
# Forced case
171
use_flashmla = (envs.VLLM_ATTENTION_BACKEND == "FLASHMLA")
172
use_cutlass_mla = (
173
- envs.VLLM_ATTENTION_BACKEND == "CUTLASS_MLA_VLLM_V1")
173
+ envs.VLLM_ATTENTION_BACKEND == "CUTLASS_MLA")
174
175
from vllm.attention.ops.flashmla import is_flashmla_supported
176
if use_flashmla and is_flashmla_supported()[0] \
@@ -182,7 +182,7 @@ class CudaPlatformBase(Platform):
182
if use_cutlass_mla and cache_config.block_size != 128:
183
cache_config.block_size = 128
184
logger.info("Forcing kv cache block size to 128 for "
185
- "CUTLASS_MLA_VLLM_V1 backend.")
185
+ "CUTLASS_MLA backend.")
186
187
compilation_config = vllm_config.compilation_config
188
if (envs.VLLM_ALL2ALL_BACKEND == "deepep_high_throughput"
@@ -211,9 +211,9 @@ class CudaPlatformBase(Platform):
211
kv_cache_dtype, block_size, use_v1,
212
use_mla) -> str:
213
if use_mla:
214
- # TODO(lucas): refactor to be more concise
214
+ # TODO(lucas): refactor to be more concise
215
# we should probably consider factoring out V1 here
216
- if selected_backend == _Backend.CUTLASS_MLA_VLLM_V1:
216
+ if selected_backend == _Backend.CUTLASS_MLA:
217
if use_v1:
218
logger.info_once("Using Cutlass MLA backend on V1 engine.")
219
return ("vllm.v1.attention.backends.mla."
vllm/platforms/interface.py
+1
-1
@@ -53,7 +53,7 @@ class _Backend(enum.Enum):
53
TRITON_MLA_VLLM_V1 = enum.auto()
54
FLASHMLA_VLLM_V1 = enum.auto()
55
FLASHMLA = enum.auto() # Supported by V1
56
- CUTLASS_MLA_VLLM_V1 = enum.auto()
56
+ CUTLASS_MLA = enum.auto()
57
PALLAS = enum.auto()
58
PALLAS_VLLM_V1 = enum.auto()
59
IPEX = enum.auto()
vllm/v1/attention/backends/mla/cutlass_mla.py
+1
-1
@@ -21,7 +21,7 @@ class CutlassMLABackend(MLACommonBackend):
21
22
@staticmethod
23
def get_name() -> str:
24
- return "CUTLASS_MLA_VLLM_V1"
24
+ return "CUTLASS_MLA"
25
26
@staticmethod
27
def get_impl_cls() -> type["CutlassMLAImpl"]: