[Kernel] [Triton] [AMD] Adding Triton implementations awq_dequantize and awq_gemm to support AWQ (#7386)

This commit is contained in:
rasmith
2024-08-28 15:37:47 -04:00
committed by GitHub
parent b98cc28f91
commit e5697d161c
5 changed files with 493 additions and 1 deletions
+9
View File
@@ -4,6 +4,7 @@ from typing import List, Optional, Tuple, Union
import torch
import vllm.envs as envs
from vllm._core_ext import ScalarType
from vllm.logger import init_logger
from vllm.platforms import current_platform
@@ -177,12 +178,20 @@ def advance_step(num_seqs: int, num_queries: int, block_size: int,
def awq_dequantize(qweight: torch.Tensor, scales: torch.Tensor,
zeros: torch.Tensor, split_k_iters: int, thx: int,
thy: int) -> torch.Tensor:
if envs.VLLM_USE_TRITON_AWQ:
from vllm.model_executor.layers.quantization.awq_triton import (
awq_dequantize_triton)
return awq_dequantize_triton(qweight, scales, zeros)
return torch.ops._C.awq_dequantize(qweight, scales, zeros, split_k_iters,
thx, thy)
def awq_gemm(input: torch.Tensor, qweight: torch.Tensor, qzeros: torch.Tensor,
scales: torch.Tensor, split_k_iters: int) -> torch.Tensor:
if envs.VLLM_USE_TRITON_AWQ:
from vllm.model_executor.layers.quantization.awq_triton import (
awq_gemm_triton)
return awq_gemm_triton(input, qweight, qzeros, scales, split_k_iters)
return torch.ops._C.awq_gemm(input, qweight, qzeros, scales, split_k_iters)