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 2, 12:15 UTC
[Refactor] Remove dead code (#43234)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Wentao Ye committed
May 29, 2026 at 00:29 UTC
710f07761741fbc20e33e8383f1b0e4c16e4655f
6 files changed
+1
-51
vllm/_custom_ops.py
-23
@@ -3719,29 +3719,6 @@ def matmul_mxf4_bf16_tn(
3719
return torch.ops._qutlass_C.matmul_mxf4_bf16_tn(a, b, a_sf, b_sf, alpha)
3720
3721
3722
-if hasattr(torch.ops._qutlass_C, "matmul_ada_mxf4_bf16_tn"):
3723
-
3724
- @register_fake("_qutlass_C::matmul_ada_mxf4_bf16_tn")
3725
- def _fake_matmul_ada_mxf4_bf16_tn(
3726
- a: torch.Tensor,
3727
- b: torch.Tensor,
3728
- a_sf: torch.Tensor,
3729
- b_sf: torch.Tensor,
3730
- alpha: torch.Tensor,
3731
- ):
3732
- return a.new_empty(*a.shape[:-1], b.shape[0], dtype=torch.bfloat16)
3733
-
3734
-
3735
-def matmul_ada_mxf4_bf16_tn(
3736
- a: torch.Tensor,
3737
- b: torch.Tensor,
3738
- a_sf: torch.Tensor,
3739
- b_sf: torch.Tensor,
3740
- alpha: torch.Tensor,
3741
-) -> torch.Tensor:
3742
- return torch.ops._qutlass_C.matmul_ada_mxf4_bf16_tn(a, b, a_sf, b_sf, alpha)
3743
-
3744
-
3722
if hasattr(torch.ops._qutlass_C, "fusedQuantizeMxQuest"):
3723
3724
@register_fake("_qutlass_C::fusedQuantizeMxQuest")
vllm/config/speculative.py
-10
@@ -724,16 +724,6 @@ class SpeculativeConfig:
724
"multiple times of forward on same MTP layer"
725
",which may result in lower acceptance rate"
726
)
727
- elif self.draft_model_config.hf_config.model_type in (
728
- "longcat_flash_mtp"
729
- ):
730
- self.method = "longcat_flash_mtp"
731
- if self.num_speculative_tokens > 1:
732
- logger.warning(
733
- "LongCat MTP models only have "
734
- "one layer. Might need some code changes "
735
- "to support multiple layers."
736
- )
727
elif self.method == "draft_model":
728
pass
729
else:
vllm/model_executor/layers/fused_moe/fused_moe.py
+1
-13
@@ -1518,7 +1518,6 @@ def fused_experts(
1518
def _get_config_quant_dtype(
1519
use_fp8_w8a8: bool,
1520
use_int8_w8a8: bool,
1521
- ocp_mx_scheme: str | None,
1521
) -> None | torch.dtype | str:
1522
"""
1523
Get the quantization type based on the quantization strategy flags.
@@ -1529,18 +1528,8 @@ def _get_config_quant_dtype(
1528
"""
1529
if use_fp8_w8a8:
1530
return current_platform.fp8_dtype()
1532
- elif use_int8_w8a8:
1531
+ if use_int8_w8a8:
1532
return torch.int8
1534
- elif ocp_mx_scheme == "w_mxfp4_a_mxfp4":
1535
- return "mxfp4"
1536
- elif ocp_mx_scheme in {"w_mxfp4_a_mxfp6_e3m2", "w_mxfp6_e3m2_a_mxfp6_e3m2"}:
1537
- return "mxfp6_e3m2"
1538
- elif ocp_mx_scheme in {"w_mxfp4_a_mxfp6_e2m3", "w_mxfp6_e2m3_a_mxfp6_e2m3"}:
1539
- return "mxfp6_e2m3"
1540
- elif ocp_mx_scheme in {"w_mxfp4", "w_mxfp6_e3m2", "w_mxfp6_e2m3"}:
1541
- return torch.bfloat16
1542
- elif ocp_mx_scheme in {"w_mxfp4_a_fp8", "w_mxfp6_e3m2_a_fp8", "w_mxfp6_e2m3_a_fp8"}:
1543
- return torch.float8_e4m3fn
1533
1534
return None
1535
@@ -1615,7 +1604,6 @@ def fused_experts_impl(
1604
quant_dtype = _get_config_quant_dtype(
1605
use_fp8_w8a8=use_fp8_w8a8,
1606
use_int8_w8a8=use_int8_w8a8,
1618
- ocp_mx_scheme=None,
1607
)
1608
1609
get_config_func = functools.partial(
vllm/model_executor/layers/linear.py
-3
@@ -50,9 +50,6 @@ WEIGHT_LOADER_V2_SUPPORTED = [
50
"AWQLinearMethod",
51
"AutoGPTQLinearMethod",
52
"Fp8LinearMethod",
53
- "MarlinLinearMethod",
54
- "GPTQMarlin24LinearMethod",
55
- "TPUInt8LinearMethod",
53
"FBGEMMFp8LinearMethod",
54
"ModelOptFp8LinearMethod",
55
"ModelOptFp8PcPtLinearMethod",
vllm/model_executor/layers/quantization/__init__.py
-1
@@ -46,7 +46,6 @@ QuantizationMethods = Literal[
46
QUANTIZATION_METHODS: list[str] = list(get_args(QuantizationMethods))
47
48
DEPRECATED_QUANTIZATION_METHODS = [
49
- "tpu_int8",
49
"fbgemm_fp8",
50
"fp_quant",
51
]
vllm/model_executor/models/qwen3_dflash.py
-1
@@ -509,7 +509,6 @@ class DFlashQwen3ForCausalLM(Qwen3ForCausalLM):
509
target_layer_num = vllm_config.model_config.get_num_layers(
510
vllm_config.parallel_config
511
)
512
- self.config.target_layer_count = target_layer_num
512
self.model = DFlashQwen3Model(
513
vllm_config=vllm_config,
514
prefix=maybe_prefix(prefix, "model"),