[LLM, MLLM] Job generation, result validation, and various fixes (#26)

Author: jaywonchung

.github/workflows/fmt_lint.yaml

@@ -0,0 +1,46 @@
+name: Check format and lint
+permissions:
+  contents: read
+
+on:
+  pull_request:
+    paths:
+      - '.github/workflows/fmt_lint.yaml'
+      - 'mlenergy/**'
+      - 'tests/**'
+      - 'scripts/*.py'
+      - 'scripts/lint.sh'
+      - 'pyproject.toml'
+      - 'uv.lock'
+  push:
+    paths:
+      - '.github/workflows/fmt_lint.yaml'
+      - 'mlenergy/**'
+      - 'tests/**'
+      - 'scripts/*.py'
+      - 'scripts/lint.sh'
+      - 'pyproject.toml'
+      - 'uv.lock'
+
+# Jobs initiated by previous pushes get cancelled by a new push.
+concurrency:
+  group: ${{ github.ref }}-format-lint
+  cancel-in-progress: true
+
+jobs:
+  format_lint_test:
+    if: ${{ github.event_name == 'push' || github.event.pull_request.head.repo.full_name != github.repository }}
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v5
+      - name: Setup uv
+        uses: astral-sh/setup-uv@v6
+        with:
+          enable-cache: true
+          python-version: "3.12"
+          version: "latest"
+      - name: Install dependencies
+        run: uv sync --dev
+      - name: Check format and lint
+        run: uv run bash scripts/lint.sh

configs/vllm/audio-chat/Qwen/Qwen3-Omni-30B-A3B-Instruct/B200/monolithic.config.yaml

@@ -4,3 +4,4 @@ max-model-len: 20480
 enable-expert-parallel: true
 dtype: bfloat16
 trust-remote-code: true
+api-server-count: 3

configs/vllm/audio-chat/Qwen/Qwen3-Omni-30B-A3B-Instruct/B200/monolithic.env.yaml

@@ -1 +1,2 @@
 VLLM_LOGGING_LEVEL: DEBUG
+VLLM_LOG_STATS_INTERVAL: 1.0

configs/vllm/audio-chat/Qwen/Qwen3-Omni-30B-A3B-Instruct/H100/monolithic.config.yaml

@@ -4,3 +4,4 @@ max-model-len: 20480
 enable-expert-parallel: true
 dtype: bfloat16
 trust-remote-code: true
+api-server-count: 3

configs/vllm/audio-chat/Qwen/Qwen3-Omni-30B-A3B-Instruct/H100/monolithic.env.yaml

@@ -1 +1,2 @@
 VLLM_LOGGING_LEVEL: DEBUG
+VLLM_LOG_STATS_INTERVAL: 1.0

configs/vllm/audio-chat/benchmark.yaml

@@ -2,15 +2,22 @@ command_template: |
   python -m mlenergy.llm.benchmark \
     --container-runtime {container_runtime} \
     --server-image {server_image} \
-    --max-output-tokens 4096 \
+    --max-output-tokens {max_output_tokens} \
     workload:audio-chat \
-    --workload.base-dir run/mllm \
+    --workload.base-dir {run_dir}/mllm \
     --workload.model-id {model_id} \
-    --workload.num-requests 1024 \
-    --workload.num-audios 1 \
+    --workload.num-requests {num_requests} \
+    --workload.num-audios {num_audios} \
     --workload.audio-data-dir ${{AUDIO_DATA_DIR:?AUDIO_DATA_DIR not set}} \
     --workload.gpu-model {gpu_model} \
+    --workload.num-gpus {num_gpus} \
     --workload.max-num-seqs {max_num_seqs}
 
+workload_defaults:
+  max_output_tokens: 4096
+  run_dir: run
+  num_requests: 1024
+  num_audios: 1
+
 sweep_defaults:
   - max_num_seqs: [8, 16, 32, 64, 96, 128, 192, 256, 384, 512]

configs/vllm/audio-chat/mistralai/Voxtral-Small-24B-2507/B200/monolithic.config.yaml

@@ -1,7 +1,9 @@
 async-scheduling: true
 max-num-batched-tokens: 8192
 max-model-len: 20480
-enable-expert-parallel: true
 dtype: bfloat16
 trust-remote-code: true
 tokenizer-mode: mistral
+config-format: mistral
+load-format: mistral
+api-server-count: 3

configs/vllm/audio-chat/mistralai/Voxtral-Small-24B-2507/B200/monolithic.env.yaml

@@ -1 +1,2 @@
 VLLM_LOGGING_LEVEL: DEBUG
+VLLM_LOG_STATS_INTERVAL: 1.0

configs/vllm/audio-chat/mistralai/Voxtral-Small-24B-2507/H100/monolithic.config.yaml

@@ -1,7 +1,9 @@
 async-scheduling: true
 max-num-batched-tokens: 8192
 max-model-len: 20480
-enable-expert-parallel: true
 dtype: bfloat16
 trust-remote-code: true
 tokenizer-mode: mistral
+config-format: mistral
+load-format: mistral
+api-server-count: 3

configs/vllm/audio-chat/mistralai/Voxtral-Small-24B-2507/H100/monolithic.env.yaml

@@ -1 +1,2 @@
 VLLM_LOGGING_LEVEL: DEBUG
+VLLM_LOG_STATS_INTERVAL: 1.0