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 10, 22:45 UTC
[Bugfix] Remove dead `block_quant_to_tensor_quant` function (#31294)
Co-authored-by: yurekami <yurekami@users.noreply.github.com> Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>
ゆり committed
Dec 25, 2025 at 01:22 UTC
506eb0f45454a07a21bf6d8731475be2a279277b
1 file changed
+1
-19
vllm/model_executor/layers/quantization/utils/fp8_utils.py
+1
-19
@@ -15,10 +15,7 @@ from vllm import _custom_ops as ops
15
from vllm._aiter_ops import rocm_aiter_ops
16
from vllm.logger import init_logger
17
from vllm.model_executor.layers.quantization.input_quant_fp8 import QuantFP8
18
-from vllm.model_executor.layers.quantization.utils.quant_utils import (
19
- GroupShape,
20
- group_broadcast,
21
-)
18
+from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
19
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
20
CUTLASS_BLOCK_FP8_SUPPORTED,
21
)
@@ -463,21 +460,6 @@ def input_to_float8(
460
return x_scl_sat.to(dtype).contiguous(), scale.float().reciprocal()
461
462
466
-def block_quant_to_tensor_quant(
467
- x_q_block: torch.Tensor,
468
- x_s: torch.Tensor,
469
-) -> tuple[torch.Tensor, torch.Tensor]:
470
- """This function converts block-wise quantization to tensor-wise
471
- quantization. The inputs are block-wise quantization tensor `x_q_block`,
472
- block-wise quantization scale and the block size.
473
- The outputs are tensor-wise quantization tensor and tensor-wise
474
- quantization scale. Note only float8 is supported for now.
475
- """
476
- x_dq_block = group_broadcast(x_q_block, x_s)
477
- x_q_tensor, scale = input_to_float8(x_dq_block, dtype=x_q_block.dtype)
478
- return x_q_tensor, scale
479
-
480
-
463
@triton.jit
464
def _per_token_group_quant_fp8(
465
# Pointers to inputs and output