A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 08:15 UTC
[Bugfix][CI/Build][Hardware][AMD] Shard ID parameters in AMD tests running parallel jobs (#9279)
Signed-off-by: Hissu Hyvarinen <hissu.hyvarinen@amd.com>
hissu-hyvarinen committed
Nov 4, 2024 at 21:37 UTC
5208dc7a203b210fa4462332a56c0012ab8b7a89
2 files changed
+12
-6
.buildkite/run-amd-test.sh
+6
-5
@@ -107,11 +107,12 @@ fi
107
PARALLEL_JOB_COUNT=8
108
# check if the command contains shard flag, we will run all shards in parallel because the host have 8 GPUs.
109
if [[ $commands == *"--shard-id="* ]]; then
110
+ # assign job count as the number of shards used
111
+ commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
112
for GPU in $(seq 0 $(($PARALLEL_JOB_COUNT-1))); do
111
- #replace shard arguments
112
- commands=${commands//"--shard-id= "/"--shard-id=${GPU} "}
113
- commands=${commands//"--num-shards= "/"--num-shards=${PARALLEL_JOB_COUNT} "}
114
- echo "Shard ${GPU} commands:$commands"
113
+ # assign shard-id for each shard
114
+ commands_gpu=${commands//"--shard-id= "/"--shard-id=${GPU} "}
115
+ echo "Shard ${GPU} commands:$commands_gpu"
116
docker run \
117
--device /dev/kfd --device /dev/dri \
118
--network host \
@@ -123,7 +124,7 @@ if [[ $commands == *"--shard-id="* ]]; then
124
-e HF_HOME=${HF_MOUNT} \
125
--name ${container_name}_${GPU} \
126
${image_name} \
126
- /bin/bash -c "${commands}" \
127
+ /bin/bash -c "${commands_gpu}" \
128
|& while read -r line; do echo ">>Shard $GPU: $line"; done &
129
PIDS+=($!)
130
done
tests/lora/test_minicpmv.py
+6
-1
@@ -1,8 +1,11 @@
1
from typing import List
2
3
+import pytest
4
+
5
import vllm
6
from vllm.assets.image import ImageAsset
7
from vllm.lora.request import LoRARequest
8
+from vllm.platforms import current_platform
9
10
MODEL_PATH = "openbmb/MiniCPM-Llama3-V-2_5"
11
@@ -53,6 +56,9 @@ def do_sample(llm: vllm.LLM, lora_path: str, lora_id: int) -> List[str]:
56
return generated_texts
57
58
59
+@pytest.mark.xfail(
60
+ current_platform.is_rocm(),
61
+ reason="MiniCPM-V dependency xformers incompatible with ROCm")
62
def test_minicpmv_lora(minicpmv_lora_files):
63
llm = vllm.LLM(
64
MODEL_PATH,
@@ -63,7 +69,6 @@ def test_minicpmv_lora(minicpmv_lora_files):
69
trust_remote_code=True,
70
gpu_memory_utilization=0.97 # This model is pretty big for CI gpus
71
)
66
-
72
output1 = do_sample(llm, minicpmv_lora_files, lora_id=1)
73
for i in range(len(EXPECTED_OUTPUT)):
74
assert EXPECTED_OUTPUT[i].startswith(output1[i])