Re-enable the 80 char line width limit (#3305)
This commit is contained in:
@@ -1,7 +1,7 @@
|
||||
"""Compare the outputs of a GPTQ model to a Marlin model.
|
||||
|
||||
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
|
||||
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
|
||||
@@ -14,7 +14,8 @@ Run `pytest tests/models/test_marlin.py --forked`.
|
||||
import pytest
|
||||
import torch
|
||||
from dataclasses import dataclass
|
||||
from vllm.model_executor.layers.quantization import _QUANTIZATION_CONFIG_REGISTRY
|
||||
from vllm.model_executor.layers.quantization import (
|
||||
_QUANTIZATION_CONFIG_REGISTRY)
|
||||
|
||||
capability = torch.cuda.get_device_capability()
|
||||
capability = capability[0] * 10 + capability[1]
|
||||
@@ -87,11 +88,11 @@ def test_models(
|
||||
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}\nMarlin:\t{marlin_output_str!r}"
|
||||
)
|
||||
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}\nMarlin:\t{marlin_output_str!r}"
|
||||
)
|
||||
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
|
||||
|
||||
Reference in New Issue
Block a user