iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[CI] Retire Weight Loading smoke tests (#57398)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Misha Goin
2026-09-17 16:04:03 -04:00
committed by GitHub
parent ac2f0ea82c
commit 9a5bd373cf
10 changed files with 0 additions and 253 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.buildkite/test-amd.yaml
-60
View File
@@ -3335,36 +3335,6 @@ steps:
commands:
- pytest -v -s v1/e2e/test_hybrid_chunked_prefill.py
#------------------------------------------------------ mi300 · weight_loading -------------------------------------------------------#
- label: ":amd: (MI300) Weight Loading Multi-GPU" # TBD
timeout_in_minutes: 70
mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
dind: false
agent_pool: mi300_2
num_gpus: 2
working_dir: "/vllm-workspace/tests"
source_file_dependencies:
- vllm/
- tests/weight_loading
optional: true
commands:
- bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt
- label: ":amd: (MI300) Weight Loading Multi-GPU (Large Models)" # TBD
timeout_in_minutes: 35
mirror_hardwares: [amdexperimental, amdproduction, amdgfx942nightly, amdmi300]
dind: false
agent_pool: mi300_2
num_gpus: 2
optional: true
working_dir: "/vllm-workspace/tests"
source_file_dependencies:
- vllm/
- tests/weight_loading
commands:
- bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt
#########################################################################################################################################
# #
# MI355 (gfx950) tests #
@@ -4719,36 +4689,6 @@ steps:
- if [ "$$BUILDKITE_PARALLEL_JOB" = "1" ]; then pytest -v -s "v1/e2e/spec_decode/acceptance_rates/mtp_other/test_mtp.py::test_gemma4_mtp_acceptance_lengths[False]"; fi
- if [ "$$BUILDKITE_PARALLEL_JOB" = "2" ]; then pytest -v -s "v1/e2e/spec_decode/acceptance_rates/mtp_other/test_mtp.py::test_gemma4_mtp_acceptance_lengths[True]"; fi
#------------------------------------------------------ mi355 · weight_loading -------------------------------------------------------#
- label: ":amd: (MI355) Weight Loading Multi-GPU" # TBD
timeout_in_minutes: 70
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
dind: false
agent_pool: mi355_2
num_gpus: 2
working_dir: "/vllm-workspace/tests"
source_file_dependencies:
- vllm/
- tests/weight_loading
optional: true
commands:
- bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt
- label: ":amd: (MI355) Weight Loading Multi-GPU (Large Models)" # TBD
timeout_in_minutes: 30
mirror_hardwares: [amdexperimental, amdproduction, amdgfx950nightly, amdmi355]
dind: false
agent_pool: mi355_2
working_dir: "/vllm-workspace/tests"
num_gpus: 2
optional: true
source_file_dependencies:
- vllm/
- tests/weight_loading
commands:
- bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large-amd.txt
#----------------------------------------------------------- mi355 · misc ------------------------------------------------------------#
- label: ":amd: (MI355) Regression" # TBD
.buildkite/test_areas/weight_loading.yaml
-38
View File
@@ -1,38 +0,0 @@
group: Weight Loading
depends_on:
- image-build
steps:
- label: ":nvidia: (L4) Weight Loading Multi-GPU"
key: weight-loading-multiple-gpu
timeout_in_minutes: 50
working_dir: "/vllm-workspace/tests"
device: l4
num_devices: 2
optional: true
source_file_dependencies:
- vllm/
- "!vllm/distributed/kv_transfer/"
- tests/weight_loading
commands:
- bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models.txt
mirror:
amd:
label: ":amd: (MI355) Weight Loading Multi-GPU"
dind: false
device: mi355_2
timeout_in_minutes: 70
depends_on:
- image-build-amd
commands:
- bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-amd.txt
# - label: Weight Loading Multiple GPU - Large Models # optional
# working_dir: "/vllm-workspace/tests"
# num_devices: 2
# device: a100
# optional: true
# source_file_dependencies:
# - vllm/
# - tests/weight_loading
# commands:
# - bash weight_loading/run_model_weight_loading_test.sh -c weight_loading/models-large.txt
.github/CODEOWNERS
-1
View File
@@ -110,7 +110,6 @@
/tests/entrypoints/llm/test_struct_output_generate.py @mgoin @aarnphm @arpera
/tests/v1/structured_output @mgoin @aarnphm @arpera
/tests/v1/core @WoosukKwon @robertgshaw2-redhat @njhill @ywang96 @alexm-redhat @heheda12345 @ApostaC @orozery @ivanium
/tests/weight_loading @mgoin @youkaichao @yewentao256
/tests/lora @jeejeelee
/tests/models/language/generation/test_hybrid.py @tdoublep @tomeras91
/tests/v1/kv_connector/nixl_integration @NickLucche
tests/kernels/quantization/test_rocm_compressed_tensors_w4a16.py
-1
View File
@@ -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",
],
)
tests/weight_loading/models-amd.txt
-18
View File
@@ -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
tests/weight_loading/models-large-amd.txt
-4
View File
@@ -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
tests/weight_loading/models-large.txt
-6
View File
@@ -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
tests/weight_loading/models.txt
-18
View File
@@ -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
tests/weight_loading/run_model_weight_loading_test.sh
-58
View File
@@ -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
tests/weight_loading/test_weight_loading.py
-49
View File
@@ -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