biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CORE] [QUANT] Support for GPTQModel's dynamic quantization per module override/control (#7086)

Browse Source
This commit is contained in:
Qubitium-ModelCloud
2025-02-13 01:19:43 +08:00
committed by GitHub
parent 2c2b560f48
commit 36a08630e8
8 changed files with 281 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
tests/quantization/test_lm_head.py
View File

@@ -3,7 +3,6 @@
Run `pytest tests/quantization/test_quant_lm_head_true.py --forked`.
"""
from typing import Tuple
import pytest
import torch
@@ -17,31 +16,31 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
PROMPT = "On the surface of Mars, we found"
MODELS_QUANT = [(
"LnL-AI/TinyLlama-1.1B-intermediate-step-1341k-3T-autoround-lm_head-symFalse",
True), ("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", False),
("neuralmagic/Meta-Llama-3-8B-Instruct-FP8", False)]
MODELS_QUANT = [
("ModelCloud/Qwen1.5-1.8B-Chat-GPTQ-4bits-dynamic-cfg-with-lm_head", True),
("ModelCloud/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit-10-25-2024", False),
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", False),
("neuralmagic/Meta-Llama-3-8B-Instruct-FP8", False)
]
@pytest.mark.parametrize("model_lm_head_quant", MODELS_QUANT)
@pytest.mark.parametrize("model_id, lm_head_quantized", MODELS_QUANT)
def test_lm_head(
vllm_runner,
model_lm_head_quant: Tuple[str, bool],
model_id: str,
lm_head_quantized: bool,
) -> None:
model, lm_head_quantized = model_lm_head_quant
with vllm_runner(model, dtype=torch.float16,
with vllm_runner(model_id, dtype=torch.float16,
max_model_len=2048) as vllm_model:
def check_model(model):
lm_head_layer = model.lm_head
if lm_head_quantized:
assert isinstance(lm_head_layer.linear_method,
assert isinstance(lm_head_layer.quant_method,
(GPTQLinearMethod, GPTQMarlinLinearMethod,
MarlinLinearMethod))
else:
assert isinstance(lm_head_layer.linear_method,
assert isinstance(lm_head_layer.quant_method,
UnquantizedEmbeddingMethod)
vllm_model.apply_model(check_model)