[ Misc ] Support Fp8 via llm-compressor (#6110)

Co-authored-by: Robert Shaw <rshaw@neuralmagic>
This commit is contained in:
Robert Shaw
2024-07-07 20:42:11 +00:00
committed by GitHub
co-authored by Robert Shaw
parent 333306a252
commit abfe705a02
17 changed files with 603 additions and 372 deletions
@@ -11,20 +11,16 @@ from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
set_weight_attrs)
from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig)
from vllm.model_executor.layers.quantization.utils.marlin_utils import (
GPTQ_MARLIN_MAX_PARALLEL, GPTQ_MARLIN_MIN_THREAD_K,
GPTQ_MARLIN_MIN_THREAD_N, GPTQ_MARLIN_SUPPORTED_GROUP_SIZES,
GPTQ_MARLIN_SUPPORTED_NUM_BITS, GPTQ_MARLIN_SUPPORTED_SYM,
GPTQ_MARLIN_TILE)
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
from vllm.platforms import current_platform
logger = init_logger(__name__)
GPTQ_MARLIN_TILE = 16
GPTQ_MARLIN_MIN_THREAD_N = 64
GPTQ_MARLIN_MIN_THREAD_K = 128
GPTQ_MARLIN_MAX_PARALLEL = 16
GPTQ_MARLIN_SUPPORTED_NUM_BITS = [4, 8]
GPTQ_MARLIN_SUPPORTED_GROUP_SIZES = [-1, 32, 64, 128]
GPTQ_MARLIN_SUPPORTED_SYM = [True]
# Permutations for Marlin scale shuffling
def get_scale_perms(num_bits: int):