[Kernel] Marlin Expansion: Support AutoGPTQ Models with Marlin (#3922)
Co-authored-by: alexm <alexm@neuralmagic.com> Co-authored-by: mgoin <michael@neuralmagic.com>
This commit is contained in:
Robert Shaw
93
tests/models/test_gptq_marlin.py
Normal file
93
tests/models/test_gptq_marlin.py
Normal file
@@ -0,0 +1,93 @@
|
||||
"""Compares the outputs of gptq vs gptq_marlin
|
||||
Note: GPTQ and Marlin do not have bitwise correctness.
|
||||
As a result, in this test, we just confirm that the top selected tokens of the
|
||||
Marlin/GPTQ models are in the top 3 selections of each other.
|
||||
Note: Marlin internally uses locks to synchronize the threads. This can
|
||||
result in very slight nondeterminism for Marlin. As a result, we re-run the test
|
||||
up to 3 times to see if we pass.
|
||||
Note: This test currently fails running with --forked with the following:
|
||||
RuntimeError: Cannot re-initialize CUDA in forked subprocess.
|
||||
To use CUDA with multiprocessing, you must use the 'spawn' start method
|
||||
Run `pytest tests/models/test_gptq_marlin.py`.
|
||||
"""
|
||||
import os
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from tests.models.utils import check_logprobs_close
|
||||
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
|
||||
|
||||
os.environ["TOKENIZERS_PARALLELISM"] = "true"
|
||||
|
||||
MAX_MODEL_LEN = 1024
|
||||
|
||||
capability = torch.cuda.get_device_capability()
|
||||
capability = capability[0] * 10 + capability[1]
|
||||
gptq_marlin_not_supported = (
|
||||
capability < QUANTIZATION_METHODS["gptq_marlin"].get_min_capability())
|
||||
|
||||
MODELS = [
|
||||
# act_order==False, group_size=channelwise
|
||||
("robertgshaw2/zephyr-7b-beta-channelwise-gptq", "main"),
|
||||
# act_order==False, group_size=128
|
||||
("TheBloke/Llama-2-7B-GPTQ", "main"),
|
||||
|
||||
# act_order==True, group_size=128
|
||||
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "main"),
|
||||
# act_order==True, group_size=64
|
||||
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-4bit-64g-actorder_True"),
|
||||
# act_order==True, group_size=32
|
||||
("TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ", "gptq-4bit-32g-actorder_True"),
|
||||
]
|
||||
|
||||
|
||||
@pytest.mark.flaky(reruns=2)
|
||||
@pytest.mark.skipif(gptq_marlin_not_supported,
|
||||
reason="gptq_marlin is not supported on this GPU type.")
|
||||
@pytest.mark.parametrize("model", MODELS)
|
||||
@pytest.mark.parametrize("dtype", ["half"])
|
||||
@pytest.mark.parametrize("max_tokens", [32])
|
||||
@pytest.mark.parametrize("num_logprobs", [5])
|
||||
def test_models(
|
||||
vllm_runner,
|
||||
example_prompts,
|
||||
model,
|
||||
dtype: str,
|
||||
max_tokens: int,
|
||||
num_logprobs: int,
|
||||
) -> None:
|
||||
model_name, revision = model
|
||||
|
||||
# Run marlin.
|
||||
gptq_marlin_model = vllm_runner(model_name=model_name,
|
||||
revision=revision,
|
||||
dtype=dtype,
|
||||
quantization="marlin",
|
||||
max_model_len=MAX_MODEL_LEN,
|
||||
tensor_parallel_size=1,
|
||||
disable_custom_all_reduce=True)
|
||||
|
||||
gptq_marlin_outputs = gptq_marlin_model.generate_greedy_logprobs(
|
||||
example_prompts, max_tokens, num_logprobs)
|
||||
del gptq_marlin_model
|
||||
|
||||
# Run gptq.
|
||||
gptq_model = vllm_runner(model_name=model_name,
|
||||
revision=revision,
|
||||
dtype=dtype,
|
||||
quantization="gptq",
|
||||
max_model_len=MAX_MODEL_LEN,
|
||||
tensor_parallel_size=1,
|
||||
disable_custom_all_reduce=True)
|
||||
gptq_outputs = gptq_model.generate_greedy_logprobs(example_prompts,
|
||||
max_tokens,
|
||||
num_logprobs)
|
||||
del gptq_model
|
||||
|
||||
check_logprobs_close(
|
||||
outputs_0_lst=gptq_outputs,
|
||||
outputs_1_lst=gptq_marlin_outputs,
|
||||
name_0="gptq",
|
||||
name_1="gptq_marlin",
|
||||
)
|
||||
@@ -10,12 +10,12 @@ up to 3 times to see if we pass.
|
||||
|
||||
Run `pytest tests/models/test_marlin.py`.
|
||||
"""
|
||||
|
||||
from dataclasses import dataclass
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from tests.models.utils import check_logprobs_close
|
||||
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
|
||||
|
||||
capability = torch.cuda.get_device_capability()
|
||||
@@ -55,43 +55,24 @@ def test_models(
|
||||
max_tokens: int,
|
||||
num_logprobs: int,
|
||||
) -> None:
|
||||
marlin_model = vllm_runner(model_pair.model_marlin, dtype=dtype)
|
||||
marlin_model = vllm_runner(model_pair.model_marlin,
|
||||
dtype=dtype,
|
||||
quantization="marlin")
|
||||
marlin_outputs = marlin_model.generate_greedy_logprobs(
|
||||
example_prompts, max_tokens, num_logprobs)
|
||||
|
||||
# Note: not sure why, but deleting just the model on Ada Lovelace
|
||||
# does not free the GPU memory. On Ampere, deleting the just model
|
||||
# frees the memory.
|
||||
del marlin_model
|
||||
|
||||
gptq_model = vllm_runner(model_pair.model_gptq, dtype=dtype)
|
||||
gptq_model = vllm_runner(model_pair.model_gptq,
|
||||
dtype=dtype,
|
||||
quantization="gptq")
|
||||
gptq_outputs = gptq_model.generate_greedy_logprobs(example_prompts,
|
||||
max_tokens,
|
||||
num_logprobs)
|
||||
|
||||
# Note: not sure why, but deleting just the model on Ada Lovelace
|
||||
# does not free the GPU memory. On Ampere, deleting the just model
|
||||
# frees the memory.
|
||||
del gptq_model
|
||||
|
||||
# loop through the prompts
|
||||
for prompt_idx in range(len(example_prompts)):
|
||||
gptq_output_ids, gptq_output_str, gptq_logprobs = gptq_outputs[
|
||||
prompt_idx]
|
||||
marlin_output_ids, marlin_output_str, marlin_logprobs = marlin_outputs[
|
||||
prompt_idx]
|
||||
|
||||
for idx, (gptq_output_id, marlin_output_id) in enumerate(
|
||||
zip(gptq_output_ids, marlin_output_ids)):
|
||||
# If sequence is not an exact match,
|
||||
if marlin_output_id != gptq_output_id:
|
||||
# Each predicted token must be in top 5 of the other's
|
||||
assert gptq_output_id in marlin_logprobs[idx], (
|
||||
f"Test{prompt_idx}:\nGPTQ:\t{gptq_output_str!r}\n"
|
||||
f"Marlin:\t{marlin_output_str!r}")
|
||||
assert marlin_output_id in gptq_logprobs[idx], (
|
||||
f"Test{prompt_idx}:\nGPTQ:\t{gptq_output_str!r}\n"
|
||||
f"Marlin:\t{marlin_output_str!r}")
|
||||
|
||||
# Break out since sequences will now diverge.
|
||||
break
|
||||
check_logprobs_close(
|
||||
outputs_0_lst=gptq_outputs,
|
||||
outputs_1_lst=marlin_outputs,
|
||||
name_0="gptq",
|
||||
name_1="marlin",
|
||||
)
|
||||
|
||||
29
tests/models/utils.py
Normal file
29
tests/models/utils.py
Normal file
@@ -0,0 +1,29 @@
|
||||
def check_logprobs_close(outputs_0_lst, outputs_1_lst, name_0, name_1):
|
||||
"""Compare the logprobs of two sequences generated by different models,
|
||||
which should be similar but not necessarily equal.
|
||||
"""
|
||||
# Loop through responses to each prompt.
|
||||
for prompt_idx, (outputs_0,
|
||||
outputs_1) in enumerate(zip(outputs_0_lst,
|
||||
outputs_1_lst)):
|
||||
output_ids_0, output_str_0, logprobs_0 = outputs_0
|
||||
output_ids_1, output_str_1, logprobs_1 = outputs_1
|
||||
|
||||
# Loop through generated tokens.
|
||||
for idx, (output_id_0,
|
||||
output_id_1) in enumerate(zip(output_ids_0, output_ids_1)):
|
||||
|
||||
# If generated tokens don't match, then
|
||||
if output_id_0 != output_id_1:
|
||||
# Each predicted token must be in top N logprobs of the other
|
||||
assert output_id_0 in logprobs_1[idx], (
|
||||
f"Test{prompt_idx}:"
|
||||
f"\n{name_0}:\t{output_str_0!r}"
|
||||
f"\n{name_1}:\t{output_str_1!r}")
|
||||
assert output_id_1 in logprobs_0[idx], (
|
||||
f"Test{prompt_idx}:"
|
||||
f"\n{name_0}:\t{output_str_0!r}"
|
||||
f"\n{name_1}:\t{output_str_1!r}")
|
||||
|
||||
# Break out since sequences will now diverge.
|
||||
break
|
||||
Reference in New Issue
Block a user