[CI] Retire Weight Loading smoke tests (#57398)
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -17,7 +17,6 @@ from vllm.platforms import current_platform
|
||||
@pytest.mark.parametrize(
|
||||
"model_path",
|
||||
[
|
||||
# Listed in tests/weight_loading/models.txt
|
||||
"nm-testing/tinyllama-oneshot-w4a16-group128-v2",
|
||||
],
|
||||
)
|
||||
|
||||
@@ -1,18 +0,0 @@
|
||||
# quantization, model, revision[, minimum capability[, safetensors load strategy]]
|
||||
fp8, amd/Meta-Llama-3.1-8B-Instruct-FP8-KV, main
|
||||
None, amd/Llama-3.2-1B-Instruct-FP8-KV, main
|
||||
fp8, amd/Mixtral-8x7B-Instruct-v0.1-FP8-KV, main, 80, prefetch
|
||||
gptq, robertgshaw2/zephyr-7b-beta-channelwise-gptq, main
|
||||
gptq, TheBloke/Llama-2-7B-GPTQ, main
|
||||
gptq, TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ, gptq-8bit--1g-actorder_True
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8-channel-a8-tensor, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8a8-dynamic-token-v2, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w4a16-group128-v2, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8a16-per-channel, main
|
||||
compressed-tensors, nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test, main
|
||||
compressed-tensors, nm-testing/Phi-3-mini-128k-instruct-FP8, main
|
||||
compressed-tensors, neuralmagic/Phi-3-medium-128k-instruct-quantized.w4a16, main
|
||||
awq, casperhansen/mixtral-instruct-awq, main
|
||||
fp8, neuralmagic/Meta-Llama-3-8B-Instruct-FP8-KV, main
|
||||
@@ -1,4 +0,0 @@
|
||||
# quantization, model, revision[, minimum capability[, safetensors load strategy]]
|
||||
fp8, amd/Meta-Llama-3.1-70B-Instruct-FP8-KV, main, 80, prefetch
|
||||
None, microsoft/phi-4, main, 80, prefetch
|
||||
fp8, amd/Mixtral-8x22B-Instruct-v0.1-FP8-KV, main, 80, prefetch
|
||||
@@ -1,6 +0,0 @@
|
||||
compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W4A16-quantized, main
|
||||
compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W4A16-channel-quantized, main
|
||||
compressed-tensors, nm-testing/Mixtral-8x7B-Instruct-v0.1-W8A16-quantized, main
|
||||
gptq, TheBloke/Mixtral-8x7B-v0.1-GPTQ, main
|
||||
awq_marlin, casperhansen/deepseek-coder-v2-instruct-awq, main
|
||||
compressed-tensors, RedHatAI/Llama-4-Scout-17B-16E-Instruct-quantized.w4a16, main
|
||||
@@ -1,18 +0,0 @@
|
||||
gptq, robertgshaw2/zephyr-7b-beta-channelwise-gptq, main
|
||||
gptq, TheBloke/Llama-2-7B-GPTQ, main
|
||||
gptq, LnL-AI/TinyLlama-1.1B-Chat-v1.0-GPTQ-4bit, main
|
||||
gptq, TheBloke/TinyLlama-1.1B-Chat-v1.0-GPTQ, gptq-8bit--1g-actorder_True
|
||||
gptq, TechxGenus/gemma-2b-it-GPTQ, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8w8-test-static-shape-change, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8-channel-a8-tensor, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8a8-dynamic-token-v2, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8a8-channel-dynamic-token-v2, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w4a16-group128-v2, main
|
||||
compressed-tensors, nm-testing/tinyllama-oneshot-w8a16-per-channel, main
|
||||
compressed-tensors, nm-testing/Meta-Llama-3-8B-FP8-compressed-tensors-test, main
|
||||
compressed-tensors, nm-testing/Phi-3-mini-128k-instruct-FP8, main
|
||||
compressed-tensors, neuralmagic/Phi-3-medium-128k-instruct-quantized.w4a16, main
|
||||
#compressed-tensors, mgoin/DeepSeek-Coder-V2-Lite-Instruct-FP8, main
|
||||
awq, casperhansen/mixtral-instruct-awq, main
|
||||
awq_marlin, casperhansen/mixtral-instruct-awq, main
|
||||
fp8, neuralmagic/Meta-Llama-3-8B-Instruct-FP8-KV, main
|
||||
@@ -1,58 +0,0 @@
|
||||
#!/bin/bash
|
||||
SUCCESS=0
|
||||
|
||||
while getopts "c:" OPT; do
|
||||
case ${OPT} in
|
||||
c )
|
||||
CONFIG="$OPTARG"
|
||||
;;
|
||||
\? )
|
||||
usage
|
||||
exit 1
|
||||
;;
|
||||
esac
|
||||
done
|
||||
|
||||
|
||||
IFS=$'\n' read -d '' -r -a MODEL_CONFIGS < "$CONFIG"
|
||||
|
||||
for MODEL_CONFIG in "${MODEL_CONFIGS[@]}"
|
||||
do
|
||||
if [[ $MODEL_CONFIG == \#* ]]; then
|
||||
echo "=== SKIPPING MODEL: $MODEL_CONFIG ==="
|
||||
continue
|
||||
fi
|
||||
|
||||
LOCAL_SUCCESS=0
|
||||
IFS=', ' read -r -a array <<< "$MODEL_CONFIG"
|
||||
|
||||
echo "=== RUNNING MODEL: $MODEL_CONFIG ==="
|
||||
|
||||
unset WEIGHT_LOADING_TEST_SAFETENSORS_LOAD_STRATEGY
|
||||
export QUANTIZATION=${array[0]}
|
||||
export MODEL_NAME=${array[1]}
|
||||
export REVISION=${array[2]}
|
||||
# If array length is larger than 3, then MIN_CAPABILITY is provided
|
||||
if [ ${#array[@]} -gt 3 ]; then
|
||||
export MIN_CAPABILITY=${array[3]}
|
||||
fi
|
||||
if [ ${#array[@]} -gt 4 ]; then
|
||||
export WEIGHT_LOADING_TEST_SAFETENSORS_LOAD_STRATEGY=${array[4]}
|
||||
fi
|
||||
pytest -s weight_loading/test_weight_loading.py || LOCAL_SUCCESS=$?
|
||||
|
||||
if [[ $LOCAL_SUCCESS == 0 ]]; then
|
||||
echo "=== PASSED MODEL: ${MODEL_CONFIG} ==="
|
||||
else
|
||||
echo "=== FAILED MODEL: ${MODEL_CONFIG} ==="
|
||||
fi
|
||||
|
||||
SUCCESS=$((SUCCESS + LOCAL_SUCCESS))
|
||||
|
||||
done
|
||||
|
||||
if [ "${SUCCESS}" -eq "0" ]; then
|
||||
exit 0
|
||||
else
|
||||
exit 1
|
||||
fi
|
||||
@@ -1,49 +0,0 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import os
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.utils.mem_constants import GiB_bytes
|
||||
|
||||
MAX_MODEL_LEN = 1024
|
||||
MODEL_NAME = os.environ.get(
|
||||
"MODEL_NAME", "robertgshaw2/zephyr-7b-beta-channelwise-gptq"
|
||||
)
|
||||
REVISION = os.environ.get("REVISION", "main")
|
||||
QUANTIZATION = os.environ.get("QUANTIZATION", "gptq_marlin")
|
||||
MIN_CAPABILITY = os.environ.get("MIN_CAPABILITY", "80")
|
||||
SAFETENSORS_LOAD_STRATEGY = os.environ.get(
|
||||
"WEIGHT_LOADING_TEST_SAFETENSORS_LOAD_STRATEGY"
|
||||
)
|
||||
|
||||
|
||||
@pytest.mark.skipif(
|
||||
MODEL_NAME == "casperhansen/deepseek-coder-v2-instruct-awq", reason="OOM in the CI"
|
||||
)
|
||||
@pytest.mark.skipif(
|
||||
not current_platform.has_device_capability(int(MIN_CAPABILITY)),
|
||||
reason="Current system does not have minimum capability.",
|
||||
)
|
||||
def test_weight_loading(vllm_runner):
|
||||
"""Test parameter weight loading with tp>1."""
|
||||
# MoE models need fp16.
|
||||
NEEDS_FP16 = QUANTIZATION == "gptq"
|
||||
with vllm_runner(
|
||||
model_name=MODEL_NAME,
|
||||
revision=REVISION,
|
||||
dtype=torch.half if NEEDS_FP16 else "auto",
|
||||
quantization=None if QUANTIZATION == "None" else QUANTIZATION,
|
||||
max_model_len=MAX_MODEL_LEN,
|
||||
tensor_parallel_size=2,
|
||||
safetensors_load_strategy=SAFETENSORS_LOAD_STRATEGY,
|
||||
# Generating 20 tokens needs a tiny KV cache, while sizing and
|
||||
# allocating a full-size one dominates this test on large devices.
|
||||
kv_cache_memory_bytes=2 * GiB_bytes,
|
||||
) as model:
|
||||
output = model.generate_greedy("Hello world!", max_tokens=20)
|
||||
print(output)
|
||||
assert output
|
||||
Reference in New Issue
Block a user