Merge pull request #321 from runpod-workers/revert-318-feat/0.24.0

Revert "feat: update to 0.24.0"

.github/workflows/CI-readme_version_sync.yml

@@ -0,0 +1,71 @@
+name: CI | Sync vLLM version in READMEs
+
+on:
+  push:
+    branches: ["main"]
+    paths:
+      - "Dockerfile"
+
+  workflow_dispatch:
+
+permissions:
+  contents: write
+  pull-requests: write
+
+jobs:
+  sync_version:
+    runs-on: ubuntu-latest
+    name: Check README version matches Dockerfile and update if needed
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Extract vLLM version from Dockerfile and sync READMEs
+        run: |
+          echo "Extracting vLLM version from Dockerfile..."
+          dockerfile_version=$(grep -oP 'vllm(?:\[[\w,]+\])?==\K[\d.]+' Dockerfile | head -1)
+
+          if [ -z "$dockerfile_version" ]; then
+            echo "ERROR: Could not extract vLLM version from Dockerfile."
+            exit 1
+          fi
+          echo "Dockerfile vLLM version: $dockerfile_version"
+          echo "VLLM_VERSION=$dockerfile_version" >> $GITHUB_ENV
+
+          updated=0
+          for readme in README.md .runpod/README.md; do
+            if [ ! -f "$readme" ]; then
+              echo "Skipping $readme (not found)"
+              continue
+            fi
+
+            readme_version=$(grep -oP 'Current vLLM version: \[\K[\d.]+' "$readme" || echo "")
+            echo "$readme current version: ${readme_version:-not found}"
+
+            if [ "$readme_version" = "$dockerfile_version" ]; then
+              echo "$readme is already up to date."
+              continue
+            fi
+
+            echo "Updating $readme from $readme_version to $dockerfile_version..."
+            sed -i "s|Current vLLM version: \[${readme_version}\](https://github.com/vllm-project/vllm/releases/tag/v${readme_version})|Current vLLM version: [${dockerfile_version}](https://github.com/vllm-project/vllm/releases/tag/v${dockerfile_version})|g" "$readme"
+            updated=1
+          done
+
+          echo "UPDATED=$updated" >> $GITHUB_ENV
+
+      - name: Create Pull Request
+        if: env.UPDATED == '1'
+        uses: peter-evans/create-pull-request@v7
+        with:
+          token: ${{ secrets.GITHUB_TOKEN }}
+          commit-message: "docs: sync vLLM version to ${{ env.VLLM_VERSION }} in READMEs"
+          title: "docs: sync vLLM version to ${{ env.VLLM_VERSION }} in READMEs"
+          body: |
+            The vLLM version in the Dockerfile has been updated to `${{ env.VLLM_VERSION }}`.
+
+            This PR syncs the version badge/link in:
+            - `README.md`
+            - `.runpod/README.md`
+          branch: docs/sync-vllm-version-${{ env.VLLM_VERSION }}
+          labels: documentation

.github/workflows/serverless-model-tests.yml

@@ -0,0 +1,162 @@
+name: Serverless Model Tests
+
+on:
+  pull_request:
+    branches:
+      - "**"
+  push:
+    branches:
+      - "main"
+
+permissions:
+  contents: read
+
+# Only one run per ref at a time — these tests spin up real H100 endpoints.
+concurrency:
+  group: serverless-model-tests-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  # On PRs we only smoke-test one (small, cheap) model to keep review loops fast.
+  test-pr:
+    if: github.event_name == 'pull_request'
+    runs-on: [blacksmith-8vcpu-ubuntu-2204, linux]
+    timeout-minutes: 60
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install script dependencies
+        run: pip install pyyaml
+
+      - name: Run serverless e2e test
+        id: e2e_test
+        run: python scripts/serverless_e2e_test.py --config configs/gpt-oss/gpt_oss_120b.yaml --build
+        env:
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY_2 }}
+          DOCKERHUB_REPO: ${{ vars.DOCKERHUB_REPO || 'runpod' }}
+          DOCKERHUB_IMG: ${{ vars.DOCKERHUB_IMG || 'worker-v1-vllm' }}
+          HUGGINGFACE_ACCESS_TOKEN: ${{ secrets.HUGGINGFACE_ACCESS_TOKEN_2 }}
+
+      - name: Cleanup safety net
+        if: always()
+        run: |
+          if [ -n "${{ steps.e2e_test.outputs.endpoint_id }}" ]; then
+            curl -sf -X DELETE "https://rest.runpod.io/v1/endpoints/${{ steps.e2e_test.outputs.endpoint_id }}" \
+              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
+          fi
+          if [ -n "${{ steps.e2e_test.outputs.template_id }}" ]; then
+            curl -sf -X DELETE "https://rest.runpod.io/v1/templates/${{ steps.e2e_test.outputs.template_id }}" \
+              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
+          fi
+
+  # On push to main we test every tuned model config in parallel.
+  discover:
+    if: github.event_name == 'push'
+    runs-on: ubuntu-latest
+    outputs:
+      configs: ${{ steps.list.outputs.configs }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: List model configs
+        id: list
+        run: |
+          CONFIGS=$(find configs -name '*.yaml' | sort | jq -R -s -c 'split("\n") | map(select(length > 0))')
+          echo "configs=$CONFIGS" >> "$GITHUB_OUTPUT"
+
+  # The image is the same regardless of which model config is under test (configs
+  # only supply endpoint env vars), so build/push it once and let every matrix job
+  # in test-main reuse that same tag instead of rebuilding per config.
+  build:
+    if: github.event_name == 'push'
+    runs-on: [blacksmith-8vcpu-ubuntu-2204, linux]
+    outputs:
+      release_version: ${{ steps.build.outputs.release_version }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up QEMU
+        uses: docker/setup-qemu-action@v3
+
+      - name: Set up Docker Buildx
+        uses: docker/setup-buildx-action@v3
+
+      - name: Login to Docker Hub
+        uses: docker/login-action@v3
+        with:
+          username: ${{ secrets.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+
+      - name: Build and push image
+        id: build
+        env:
+          DOCKERHUB_REPO: ${{ vars.DOCKERHUB_REPO || 'runpod' }}
+          DOCKERHUB_IMG: ${{ vars.DOCKERHUB_IMG || 'worker-v1-vllm' }}
+          RELEASE_VERSION: test-${{ github.sha }}
+          HUGGINGFACE_ACCESS_TOKEN: ${{ secrets.HUGGINGFACE_ACCESS_TOKEN }}
+        run: |
+          docker buildx bake --push
+          echo "release_version=${RELEASE_VERSION}" >> "$GITHUB_OUTPUT"
+
+  test-main:
+    needs: [discover, build]
+    if: github.event_name == 'push' && needs.discover.result == 'success' && needs.build.result == 'success'
+    runs-on: [blacksmith-8vcpu-ubuntu-2204, linux]
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        config: ${{ fromJSON(needs.discover.outputs.configs) }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install script dependencies
+        run: pip install pyyaml
+
+      - name: Run serverless e2e test
+        id: e2e_test
+        run: python scripts/serverless_e2e_test.py --config "${{ matrix.config }}" --image "${DOCKERHUB_REPO}/${DOCKERHUB_IMG}:${RELEASE_VERSION}"
+        env:
+          RUNPOD_API_KEY: ${{ secrets.RUNPOD_API_KEY_2 }}
+          HUGGINGFACE_ACCESS_TOKEN: ${{ secrets.HUGGINGFACE_ACCESS_TOKEN_2 }}
+          DOCKERHUB_REPO: ${{ vars.DOCKERHUB_REPO || 'runpod' }}
+          DOCKERHUB_IMG: ${{ vars.DOCKERHUB_IMG || 'worker-v1-vllm' }}
+          RELEASE_VERSION: ${{ needs.build.outputs.release_version }}
+
+      - name: Cleanup safety net
+        if: always()
+        run: |
+          if [ -n "${{ steps.e2e_test.outputs.endpoint_id }}" ]; then
+            curl -sf -X DELETE "https://rest.runpod.io/v1/endpoints/${{ steps.e2e_test.outputs.endpoint_id }}" \
+              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
+          fi
+          if [ -n "${{ steps.e2e_test.outputs.template_id }}" ]; then
+            curl -sf -X DELETE "https://rest.runpod.io/v1/templates/${{ steps.e2e_test.outputs.template_id }}" \
+              -H "Authorization: Bearer ${{ secrets.RUNPOD_API_KEY_2 }}" || true
+          fi

.github/workflows/test.yml

@@ -0,0 +1,32 @@
+name: Tests
+
+on:
+  pull_request:
+    branches:
+      - "**"
+  push:
+    branches:
+      - "main"
+
+permissions:
+  contents: read
+
+jobs:
+  pytest:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install test dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r tests/requirements.txt
+
+      - name: Run unit tests
+        run: python -m pytest tests -v

.runpod/README.md

@@ -6,7 +6,7 @@ Run LLMs using [vLLM](https://docs.vllm.ai) with an OpenAI-compatible API
 
 [![RunPod](https://api.runpod.io/badge/runpod-workers/worker-vllm)](https://www.runpod.io/console/hub/runpod-workers/worker-vllm)
 
-Current vLLM version: [0.20.2](https://github.com/vllm-project/vllm/releases/tag/v0.20.2)
+Current vLLM version: [0.23.0](https://github.com/vllm-project/vllm/releases/tag/v0.23.0)
 
 ---
 
@@ -33,6 +33,17 @@ All behaviour is controlled through environment variables:
 
 **Pass any vLLM engine arg** not listed above by setting an env var with the **UPPERCASED** field name (e.g. `MAX_MODEL_LEN=4096`, `ENABLE_CHUNKED_PREFILL=true`). The worker auto-discovers all `AsyncEngineArgs` fields from env. See the [vLLM engine args docs](https://docs.vllm.ai/en/latest/configuration/engine_args) for all available options.
 
+**Configuration file:** You can also supply a `config.yaml` instead of (or alongside) env vars. Mount it at `/vllm_config.yaml` in the container, or set `VLLM_CONFIG_FILE` to a custom path. Use the same key names as `vllm serve` — hyphens and underscores both work:
+
+```yaml
+model: meta-llama/Llama-3.1-8B-Instruct
+max-model-len: 8192
+gpu-memory-utilization: 0.90
+quantization: awq
+```
+
+Environment variables always override config file values.
+
 For complete configuration options, see the [full configuration documentation](https://github.com/runpod-workers/worker-vllm/blob/main/docs/configuration.md).
 
 ### Specify Transformers Version

.runpod/hub.json

@@ -247,6 +247,7 @@
           "name": "Max Parallel Loading Workers",
           "type": "number",
           "description": "Load model sequentially in multiple batches.",
+          "default": 1,
           "advanced": true
         }
       },

.runpod/tests.json

@@ -30,7 +30,7 @@
     }
   ],
   "config": {
-    "gpuTypeId": "NVIDIA GeForce RTX 4090",
+    "gpuTypeId": "NVIDIA L40",
     "gpuCount": 1,
     "env": [
       {

Dockerfile

@@ -1,17 +1,36 @@
 FROM nvidia/cuda:13.0.2-devel-ubuntu22.04
 
+ENV DEBIAN_FRONTEND=noninteractive
+
 RUN apt-get update -y \
-    && apt-get install -y python3-pip curl git \
+    && apt-get install -y curl git software-properties-common \
-    && curl -LsSf https://astral.sh/uv/install.sh  | sh
+    && add-apt-repository -y ppa:deadsnakes/ppa \
+    && apt-get install -y python3.12 python3.12-dev python3.12-venv \
+    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.12 1 \
+    && update-alternatives --set python3 /usr/bin/python3.12 \
+    && rm -f /usr/lib/python3.12/EXTERNALLY-MANAGED \
+    && curl -LsSf https://astral.sh/uv/install.sh | sh
 
 ENV PATH="/root/.local/bin:$PATH"
 
 RUN ldconfig /usr/local/cuda-13.0/compat/
 
 # Install vLLM with FlashInfer - use CUDA 130 PyTorch wheels
-RUN uv pip install --system "packaging>=24.2" && \
+RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system "vllm[flashinfer]==0.20.2" && \
+    uv pip install --system "packaging>=24.2" && \
-    uv pip install --system git+https://github.com/deepseek-ai/DeepGEMM.git@714dd1a4a980f7937a74343d19a8eba4fe321480 --no-build-isolation
+    uv pip install --system "vllm[flashinfer]==0.23.0" && \
+    uv pip install --system git+https://github.com/deepseek-ai/DeepGEMM.git@714dd1a4a980f7937a74343d19a8eba4fe321480 --no-build-isolation && \
+    uv pip install --system --force-reinstall --no-deps nixl-cu13
+
+# Fix CUTLASS DSL cu13 install order: nvidia-cutlass-dsl[cu13] installs
+# -libs-base and -libs-cu13 wheels that share paths with different content.
+# uv can extract them in either order, leaving base files that break CUDA 13
+# CuTe DSL JIT. Force -libs-cu13 last. See vllm-project/vllm#45204.
+RUN CUTLASS_DSL_VERSION=$(uv pip show --system nvidia-cutlass-dsl 2>/dev/null | awk '/^Version:/{print $2}') && \
+    if [ -n "$CUTLASS_DSL_VERSION" ]; then \
+        uv pip install --system --force-reinstall --no-deps \
+            "nvidia-cutlass-dsl-libs-cu13==${CUTLASS_DSL_VERSION}"; \
+    fi
 
 # Install additional Python dependencies (after vLLM to avoid PyTorch version conflicts)
 COPY builder/requirements.txt /requirements.txt
@@ -47,7 +66,8 @@ ENV MODEL_NAME=$MODEL_NAME \
     # Disable DeepGEMM MoE kernels by default; override with VLLM_USE_DEEP_GEMM=1 to enable
     VLLM_USE_DEEP_GEMM=0
 
-ENV PYTHONPATH="/:/vllm-workspace"
+ENV PYTHONPATH="/:/vllm-workspace" \
+    LD_LIBRARY_PATH="/usr/local/nvidia/lib64:/usr/local/cuda/lib64:${LD_LIBRARY_PATH}"
 
 RUN if [ "${VLLM_NIGHTLY}" = "true" ]; then \
     uv pip install --system -U vllm --pre --index-url https://pypi.org/simple --extra-index-url https://wheels.vllm.ai/nightly && \

README.md

@@ -8,7 +8,7 @@ Deploy OpenAI-Compatible Blazing-Fast LLM Endpoints powered by the [vLLM](https:
 
 ![vLLM worker banner](https://image.runpod.ai/preview/vllm/vllm-banner.png)
 
-Current vLLM version: [0.20.2](https://github.com/vllm-project/vllm/releases/tag/v0.20.2)
+Current vLLM version: [0.23.0](https://github.com/vllm-project/vllm/releases/tag/v0.23.0)
 
 
 > Check out our Load Balancer implementation here: [vLLM Load Balancer](https://github.com/runpod-workers/vllm-loadbalancer-ep)
@@ -78,6 +78,20 @@ Configure worker-vllm using environment variables:
 
 Any env var whose name matches a valid `AsyncEngineArgs` field (uppercased) is applied automatically. Backward-compat aliases: `MODEL_NAME`, `TOKENIZER_NAME`, `MAX_CONTEXT_LEN_TO_CAPTURE`. This lets you configure any vLLM option without waiting for explicit worker support.
 
+### Configuration File (config.yaml)
+
+As an alternative to environment variables, you can supply a `config.yaml` file using the same key names as `vllm serve` (hyphens or underscores both work):
+
+```yaml
+model: meta-llama/Llama-3.1-8B-Instruct
+max-model-len: 8192
+gpu-memory-utilization: 0.90
+quantization: awq
+tensor-parallel-size: 2
+```
+
+Mount the file into the container at `/vllm_config.yaml`, or point to a custom path with the `VLLM_CONFIG_FILE` env var. Environment variables always take precedence over config file values.
+
 For the complete list of all available environment variables, examples, and detailed descriptions: **[Configuration](docs/configuration.md)**
 
 ### Specify Transformers Version

builder/requirements.txt

@@ -1,9 +1,9 @@
 ray
 pandas
 pyarrow
-runpod==1.9.0
+runpod~=1.11.0
 huggingface-hub
-lmcache==0.4.5
+lmcache==0.5.0
 packaging>=24.2
 typing-extensions>=4.8.0
 pydantic

configs/gemma/gemma4_31b_it.yaml

@@ -0,0 +1,12 @@
+model: google/gemma-4-31b-it
+gpu-memory-utilization: 0.95
+max-model-len: 8192
+dtype: auto
+trust-remote-code: true
+quantization: fp8
+kv-cache-dtype: fp8
+enforce-eager: false
+enable-prefix-caching: true
+enable-chunked-prefill: true
+vllm-release: v2.22.5 
+speculative-config: '{"model":"RedHatAI/gemma-4-31B-it-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'

configs/gpt-oss/gpt_oss_120b.yaml

@@ -0,0 +1,9 @@
+model: openai/gpt-oss-120b
+gpu-memory-utilization: 0.95
+max-model-len: 8192
+dtype: auto
+trust-remote-code: true
+enforce-eager: false
+enable-prefix-caching: true
+enable-chunked-prefill: true
+speculative-config: '{"model":"RedHatAI/gpt-oss-120b-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'

configs/llama/llama3.1_8b_instruct.yaml

@@ -0,0 +1,11 @@
+model: meta-llama/Llama-3.1-8B-Instruct
+gpu-memory-utilization: 0.95
+max-model-len: 8192
+dtype: auto
+trust-remote-code: true
+quantization: fp8
+kv-cache-dtype: fp8
+enforce-eager: false
+vllm-release: v2.22.5 
+enable-prefix-caching: true
+speculative-config: '{"model":"RedHatAI/Llama-3.1-8B-Instruct-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'

configs/qwen/qwen3_8b.yaml

@@ -0,0 +1,12 @@
+model: Qwen/Qwen3-8B
+gpu-memory-utilization: 0.95
+max-model-len: 8192
+dtype: auto
+trust-remote-code: true
+quantization: fp8
+kv-cache-dtype: fp8
+enforce-eager: false
+enable-prefix-caching: true
+vllm-release: v2.22.5 
+compilation-config: '{"cudagraph_mode": "PIECEWISE"}'
+speculative-config: '{"model":"RedHatAI/Qwen3-8B-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'

scripts/serverless_e2e_test.py

@@ -0,0 +1,328 @@
+#!/usr/bin/env python3
+"""End-to-end test: build/push the worker image, deploy it as a real RunPod
+Serverless endpoint on H100, run the .runpod/tests.json test cases against it,
+then tear the endpoint down.
+
+Usage:
+    RUNPOD_API_KEY=... python scripts/serverless_e2e_test.py \\
+        --config configs/qwen/qwen3_8b.yaml --build
+
+    RUNPOD_API_KEY=... python scripts/serverless_e2e_test.py \\
+        --config configs/qwen/qwen3_8b.yaml --image runpod/worker-v1-vllm:dev-my-branch
+"""
+import argparse
+import json
+import os
+import subprocess
+import sys
+import time
+import urllib.error
+import urllib.request
+import uuid
+from pathlib import Path
+
+import yaml
+
+REPO_ROOT = Path(__file__).resolve().parent.parent
+
+REST_API_BASE = "https://rest.runpod.io/v1"
+JOB_API_BASE = "https://api.runpod.ai/v2"
+
+DEFAULT_GPU_TYPE_IDS = [
+    "NVIDIA H100 80GB HBM3",
+    "NVIDIA H100 NVL",
+    "NVIDIA H100 PCIe",
+]
+
+TERMINAL_STATUSES = {"COMPLETED", "FAILED", "TIMED_OUT", "CANCELLED"}
+
+
+def log(msg: str) -> None:
+    print(f"[serverless_e2e_test] {msg}", flush=True)
+
+
+def write_github_output(key: str, value: str) -> None:
+    """Expose a value to later CI steps (e.g. an always() cleanup safety net
+    for when this process gets killed before its own `finally` can run)."""
+    path = os.environ.get("GITHUB_OUTPUT")
+    if not path:
+        return
+    with open(path, "a") as f:
+        f.write(f"{key}={value}\n")
+
+
+def stringify_env_value(value) -> str:
+    if isinstance(value, bool):
+        return "true" if value else "false"
+    if isinstance(value, (dict, list)):
+        return json.dumps(value)
+    return str(value)
+
+
+def load_hub_defaults(hub_json_path: Path) -> dict:
+    hub = json.loads(hub_json_path.read_text())
+    config = hub.get("config", {})
+
+    base_env = {}
+    for entry in config.get("env", []):
+        default = entry.get("input", {}).get("default")
+        if default is None:
+            continue
+        base_env[entry["key"]] = stringify_env_value(default)
+
+    return {
+        "containerDiskInGb": config.get("containerDiskInGb", 50),
+        "gpuCount": config.get("gpuCount", 1),
+        "allowedCudaVersions": config.get("allowedCudaVersions"),
+        "minCudaVersion": config.get("minCudaVersion"),
+        "env": base_env,
+    }
+
+
+
+# config keys whose naive `KEY.replace("-", "_").upper()` transform doesn't match
+# the env var the worker actually reads (see src/engine_args.py ENV_ALIASES).
+CONFIG_KEY_ALIASES = {
+    "MODEL": "MODEL_NAME",
+}
+
+
+def load_model_env(config_yaml_path: Path) -> dict:
+    raw = yaml.safe_load(config_yaml_path.read_text()) or {}
+    env = {}
+    for key, value in raw.items():
+        env_key = key.replace("-", "_").upper()
+        env_key = CONFIG_KEY_ALIASES.get(env_key, env_key)
+        env[env_key] = stringify_env_value(value)
+    return env
+
+
+def build_and_push_image(dockerhub_repo: str, dockerhub_img: str, release_version: str) -> str:
+    hf_token = os.environ.get("HUGGINGFACE_ACCESS_TOKEN", "")
+    dockerhub_user = os.environ.get("DOCKERHUB_USERNAME")
+    dockerhub_pass = os.environ.get("DOCKERHUB_TOKEN")
+
+    if dockerhub_user and dockerhub_pass:
+        log(f"Logging in to Docker Hub as {dockerhub_user}")
+        subprocess.run(
+            ["docker", "login", "-u", dockerhub_user, "--password-stdin"],
+            input=dockerhub_pass,
+            text=True,
+            check=True,
+            cwd=REPO_ROOT,
+        )
+    else:
+        log("DOCKERHUB_USERNAME/DOCKERHUB_TOKEN not set; assuming docker is already logged in")
+
+    tag = f"{dockerhub_repo}/{dockerhub_img}:{release_version}"
+    log(f"Building and pushing {tag} via docker buildx bake")
+    # docker-bake.hcl's DOCKERHUB_REPO/DOCKERHUB_IMG/RELEASE_VERSION are bake-level
+    # `variable` blocks that compute `tags` - they read from env vars of the same
+    # name, NOT from `--set target.args.*` (that sets Dockerfile build ARGs, a
+    # separate namespace the Dockerfile doesn't even declare these under).
+    bake_env = {
+        **os.environ,
+        "DOCKERHUB_REPO": dockerhub_repo,
+        "DOCKERHUB_IMG": dockerhub_img,
+        "RELEASE_VERSION": release_version,
+        "HUGGINGFACE_ACCESS_TOKEN": hf_token,
+    }
+    subprocess.run(
+        ["docker", "buildx", "bake", "--push"],
+        check=True,
+        cwd=REPO_ROOT,
+        env=bake_env,
+    )
+    return tag
+
+
+def api_request(method: str, url: str, api_key: str, body: dict | None = None) -> dict:
+    data = json.dumps(body).encode() if body is not None else None
+    req = urllib.request.Request(url, data=data, method=method)
+    req.add_header("Authorization", f"Bearer {api_key}")
+    req.add_header("Content-Type", "application/json")
+    try:
+        with urllib.request.urlopen(req) as resp:
+            raw = resp.read()
+            return json.loads(raw) if raw else {}
+    except urllib.error.HTTPError as e:
+        detail = e.read().decode(errors="replace")
+        raise RuntimeError(f"{method} {url} -> HTTP {e.code}: {detail}") from e
+
+
+def create_template(api_key: str, name: str, image: str, env: dict, container_disk_in_gb: int) -> str:
+    log(f"Creating template {name!r} for image {image}")
+    resp = api_request("POST", f"{REST_API_BASE}/templates", api_key, {
+        "name": name,
+        "imageName": image,
+        "isServerless": True,
+        "env": env,
+        "containerDiskInGb": container_disk_in_gb,
+    })
+    return resp["id"]
+
+
+def create_endpoint(
+    api_key: str,
+    name: str,
+    template_id: str,
+    gpu_type_ids: list[str],
+    gpu_count: int,
+    allowed_cuda_versions: list[str] | None,
+    min_cuda_version: str | None,
+    idle_timeout: int,
+) -> str:
+    log(f"Creating endpoint {name!r} (gpuTypeIds={gpu_type_ids})")
+    body = {
+        "name": name,
+        "templateId": template_id,
+        "gpuTypeIds": gpu_type_ids,
+        "gpuCount": gpu_count,
+        "workersMin": 0,
+        "workersMax": 1,
+        "idleTimeout": idle_timeout,
+        "scalerType": "QUEUE_DELAY",
+        "scalerValue": 4,
+    }
+    if allowed_cuda_versions:
+        body["allowedCudaVersions"] = allowed_cuda_versions
+    if min_cuda_version:
+        body["minCudaVersion"] = min_cuda_version
+    resp = api_request("POST", f"{REST_API_BASE}/endpoints", api_key, body)
+    return resp["id"]
+
+
+def delete_endpoint(api_key: str, endpoint_id: str) -> None:
+    log(f"Deleting endpoint {endpoint_id}")
+    try:
+        api_request("DELETE", f"{REST_API_BASE}/endpoints/{endpoint_id}", api_key)
+    except RuntimeError as e:
+        log(f"WARNING: failed to delete endpoint {endpoint_id}: {e}")
+
+
+def delete_template(api_key: str, template_id: str) -> None:
+    log(f"Deleting template {template_id}")
+    try:
+        api_request("DELETE", f"{REST_API_BASE}/templates/{template_id}", api_key)
+    except RuntimeError as e:
+        log(f"WARNING: failed to delete template {template_id}: {e}")
+
+
+def response_has_error(output) -> bool:
+    if isinstance(output, dict):
+        return "error" in output
+    if isinstance(output, list):
+        return any(isinstance(item, dict) and "error" in item for item in output)
+    return False
+
+
+def run_test_case(api_key: str, endpoint_id: str, test: dict, cold_start_buffer_seconds: int) -> bool:
+    name = test.get("name", "unnamed_test")
+    deadline = time.monotonic() + test.get("timeout", 300000) / 1000 + cold_start_buffer_seconds
+
+    log(f"Submitting job for test {name!r}")
+    submit = api_request("POST", f"{JOB_API_BASE}/{endpoint_id}/run", api_key, {"input": test["input"]})
+    job_id = submit["id"]
+
+    while True:
+        if time.monotonic() > deadline:
+            log(f"FAIL {name}: timed out waiting for job {job_id}")
+            return False
+
+        status_resp = api_request("GET", f"{JOB_API_BASE}/{endpoint_id}/status/{job_id}", api_key)
+        status = status_resp.get("status")
+
+        if status in TERMINAL_STATUSES:
+            if status != "COMPLETED":
+                log(f"FAIL {name}: job {job_id} ended with status {status}: {status_resp}")
+                return False
+            if response_has_error(status_resp.get("output")):
+                log(f"FAIL {name}: job {job_id} completed but output contained an error: {status_resp.get('output')}")
+                return False
+            log(f"PASS {name}")
+            return True
+
+        time.sleep(5)
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--config", required=True, type=Path, help="Path to a configs/**/*.yaml vLLM config")
+    parser.add_argument("--image", help="Existing image tag to test; skips building unless --build is also given")
+    parser.add_argument("--build", action="store_true", help="Build and push the image before testing")
+    parser.add_argument("--keep", action="store_true", help="Don't tear down the endpoint/template afterward")
+    parser.add_argument("--hub-json", type=Path, default=REPO_ROOT / ".runpod" / "hub.json")
+    parser.add_argument("--tests-file", type=Path, default=REPO_ROOT / ".runpod" / "tests.json")
+    parser.add_argument("--gpu-type-ids", default=",".join(DEFAULT_GPU_TYPE_IDS))
+    parser.add_argument("--min-cuda-version", help="Overrides the config/hub.json minCudaVersion")
+    parser.add_argument("--dockerhub-repo", default=os.environ.get("DOCKERHUB_REPO", "runpod"))
+    parser.add_argument("--dockerhub-img", default=os.environ.get("DOCKERHUB_IMG", "worker-v1-vllm"))
+    parser.add_argument("--idle-timeout", type=int, default=60)
+    parser.add_argument("--cold-start-buffer-seconds", type=int, default=600)
+    args = parser.parse_args()
+
+    api_key = os.environ.get("RUNPOD_API_KEY")
+    if not api_key:
+        log("ERROR: RUNPOD_API_KEY is not set")
+        return 1
+
+    model_slug = args.config.stem
+    run_id = uuid.uuid4().hex[:8]
+
+    if args.build or not args.image:
+        release_version = f"test-{model_slug}-{run_id}"
+        image = build_and_push_image(args.dockerhub_repo, args.dockerhub_img, release_version)
+    else:
+        image = args.image
+
+    hub_defaults = load_hub_defaults(args.hub_json)
+    model_env = load_model_env(args.config)
+    env = {**hub_defaults["env"], **model_env}
+    hf_token = os.environ.get("HF_TOKEN") or os.environ.get("HUGGINGFACE_ACCESS_TOKEN")
+    if hf_token:
+        env["HF_TOKEN"] = hf_token
+    tests_data = json.loads(args.tests_file.read_text())
+    gpu_type_ids = [g.strip() for g in args.gpu_type_ids.split(",") if g.strip()]
+
+    resource_name = f"worker-vllm-e2e-{model_slug}-{run_id}"
+    template_id = None
+    endpoint_id = None
+    try:
+        template_id = create_template(
+            api_key, resource_name, image, env, hub_defaults["containerDiskInGb"]
+        )
+        write_github_output("template_id", template_id)
+        endpoint_id = create_endpoint(
+            api_key,
+            resource_name,
+            template_id,
+            gpu_type_ids,
+            hub_defaults["gpuCount"],
+            hub_defaults["allowedCudaVersions"],
+            args.min_cuda_version or hub_defaults["minCudaVersion"],
+            args.idle_timeout,
+        )
+        write_github_output("endpoint_id", endpoint_id)
+
+        results = [
+            run_test_case(api_key, endpoint_id, test, args.cold_start_buffer_seconds)
+            for test in tests_data["tests"]
+        ]
+
+        if all(results):
+            log(f"All {len(results)} test(s) passed for {model_slug}")
+            return 0
+        log(f"{results.count(False)}/{len(results)} test(s) failed for {model_slug}")
+        return 1
+    finally:
+        if not args.keep:
+            if endpoint_id:
+                delete_endpoint(api_key, endpoint_id)
+            if template_id:
+                delete_template(api_key, template_id)
+        else:
+            log(f"--keep passed; leaving endpoint={endpoint_id} template={template_id} running")
+
+
+if __name__ == "__main__":
+    sys.exit(main())

src/engine.py

@@ -9,7 +9,6 @@ from typing import AsyncGenerator, Optional
 from dotenv import load_dotenv
 from vllm import AsyncLLMEngine
 from vllm.inputs import TextPrompt
-from vllm.entrypoints.logger import RequestLogger
 from vllm.entrypoints.anthropic.protocol import AnthropicMessagesRequest, AnthropicMessagesResponse, AnthropicError, AnthropicErrorResponse
 from vllm.entrypoints.anthropic.serving import AnthropicServingMessages
 from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest

src/engine_args.py

@@ -404,6 +404,23 @@ def _resolve_cached_model_path(model_name: str) -> str:
     return resolved
 
 
+def _get_args_from_config_file() -> dict:
+    """Load engine args from a vLLM-style config.yaml.
+
+    Checks VLLM_CONFIG_FILE env var, then falls back to /vllm_config.yaml.
+    Keys use the same long-form names as vllm serve (hyphens converted to underscores).
+    """
+    import yaml
+    path = os.getenv("VLLM_CONFIG_FILE", "/vllm_config.yaml")
+    if not os.path.exists(path):
+        return {}
+    with open(path) as f:
+        raw = yaml.safe_load(f) or {}
+    normalized = {k.replace("-", "_"): v for k, v in raw.items()}
+    logging.info("Loaded engine args from config file %s: %s", path, list(normalized.keys()))
+    return normalized
+
+
 def get_local_args():
     """
     Retrieve local arguments from a JSON file.
@@ -429,6 +446,9 @@ def get_engine_args():
     # Start with worker custom defaults (only where we differ from vLLM)
     args = dict(DEFAULT_ARGS)
 
+    # Config file values sit above defaults but below env vars
+    args.update(_get_args_from_config_file())
+
     # Auto-discover: every AsyncEngineArgs field from env UPPERCASED (e.g. MAX_MODEL_LEN)
     args.update(_get_args_from_env_auto_discover())
 
@@ -581,6 +601,13 @@ def get_engine_args():
 
     # Resolve lowercase HF cache paths (FDE-174)
     if args.get("model"):
-        args["model"] = _resolve_cached_model_path(args["model"])
+        original_model = args["model"]
+        args["model"] = _resolve_cached_model_path(original_model)
+        # When the model was rewritten to an on-disk snapshot path, keep serving
+        # under the original repo id so the OpenAI API model name does not become
+        # a filesystem path (issue #310). An explicit served_model_name (or the
+        # OPENAI_SERVED_MODEL_NAME_OVERRIDE handled downstream) still wins.
+        if args["model"] != original_model and not args.get("served_model_name"):
+            args["served_model_name"] = original_model
 
     return AsyncEngineArgs(**args)

tests/conftest.py

@@ -0,0 +1,105 @@
+"""Shared test fixtures.
+
+``src/engine_args.py`` hard-imports ``vllm`` (and a tensorizer submodule) and
+``torch.cuda``. Both are only installed inside the GPU Docker image, so when the
+tests run on a machine without them we install lightweight stubs. When the real
+packages *are* available (e.g. CI inside the worker image) the stubs are skipped
+and the real ones are used instead.
+"""
+
+import sys
+import types
+from dataclasses import dataclass
+from typing import Optional, Union, List
+
+
+def _install_torch_stub():
+    try:
+        import torch  # noqa: F401
+        return  # real torch present, nothing to stub
+    except Exception:
+        pass
+
+    torch = types.ModuleType("torch")
+    cuda = types.ModuleType("torch.cuda")
+    # No GPU in the test environment -> 0 devices (skips tensor-parallel setup).
+    cuda.device_count = lambda: 0
+    torch.cuda = cuda
+    sys.modules["torch"] = torch
+    sys.modules["torch.cuda"] = cuda
+
+
+def _install_vllm_stub():
+    try:
+        import vllm  # noqa: F401
+        return  # real vLLM present, nothing to stub
+    except Exception:
+        pass
+
+    vllm = types.ModuleType("vllm")
+
+    @dataclass
+    class AsyncEngineArgs:
+        # Only the fields the worker actually sets/reads need to exist here;
+        # get_engine_args() filters args down to AsyncEngineArgs.__dataclass_fields__
+        # before construction, so unknown keys are dropped rather than passed.
+        model: Optional[str] = None
+        served_model_name: Optional[Union[str, List[str]]] = None
+        revision: Optional[str] = None
+        tokenizer: Optional[str] = None
+        trust_remote_code: bool = False
+        max_model_len: Optional[int] = None
+        max_num_batched_tokens: Optional[int] = None
+        disable_log_stats: bool = False
+        gpu_memory_utilization: float = 0.9
+        tensor_parallel_size: int = 1
+        max_parallel_loading_workers: Optional[int] = None
+        kv_cache_dtype: Optional[str] = None
+
+    class _Stub:  # pragma: no cover - placeholder for vllm symbols
+        def __init__(self, *args, **kwargs):
+            pass
+
+    vllm.AsyncEngineArgs = AsyncEngineArgs
+    vllm.SamplingParams = _Stub
+    sys.modules["vllm"] = vllm
+
+    # src.utils imports these at module load and uses ErrorResponse as a return
+    # annotation, which Python evaluates eagerly on <3.14 -> must be defined.
+    vllm_utils = types.ModuleType("vllm.utils")
+    vllm_utils.random_uuid = lambda: "stub-uuid"
+    vllm.utils = vllm_utils
+    sys.modules["vllm.utils"] = vllm_utils
+
+    protocol = types.ModuleType("vllm.entrypoints.openai.engine.protocol")
+    protocol.ErrorResponse = _Stub
+    protocol.ErrorInfo = _Stub
+    protocol.RequestResponseMetadata = _Stub
+    for name in (
+        "vllm.entrypoints",
+        "vllm.entrypoints.openai",
+        "vllm.entrypoints.openai.engine",
+    ):
+        sys.modules.setdefault(name, types.ModuleType(name))
+    sys.modules["vllm.entrypoints.openai.engine.protocol"] = protocol
+
+    # vllm.model_executor.model_loader.tensorizer.TensorizerConfig
+    model_executor = types.ModuleType("vllm.model_executor")
+    model_loader = types.ModuleType("vllm.model_executor.model_loader")
+    tensorizer = types.ModuleType("vllm.model_executor.model_loader.tensorizer")
+
+    class TensorizerConfig:  # pragma: no cover - placeholder
+        def __init__(self, *args, **kwargs):
+            pass
+
+    tensorizer.TensorizerConfig = TensorizerConfig
+    model_loader.tensorizer = tensorizer
+    model_executor.model_loader = model_loader
+    vllm.model_executor = model_executor
+    sys.modules["vllm.model_executor"] = model_executor
+    sys.modules["vllm.model_executor.model_loader"] = model_loader
+    sys.modules["vllm.model_executor.model_loader.tensorizer"] = tensorizer
+
+
+_install_torch_stub()
+_install_vllm_stub()

tests/requirements.txt

@@ -0,0 +1,6 @@
+# Test-only dependencies. vllm/torch are stubbed in conftest.py when absent,
+# so the unit tests run on a plain CPU runner without the GPU image.
+pytest>=8,<10
+# get_engine_args() reads a vLLM-style config via PyYAML (a transitive vllm dep
+# at runtime); install it explicitly here since vllm itself is stubbed.
+pyyaml

tests/test_engine_args.py

@@ -0,0 +1,126 @@
+"""Tests for HF cache path resolution and served-model-name decoupling.
+
+Regression coverage for issue #310: when MODEL_NAME is served from a lowercased
+HF cache dir, the cache resolver rewrites engine_args.model to a snapshot path.
+The served model name must stay the original repo id, not the path.
+"""
+
+import os
+
+import pytest
+
+from src import engine_args
+from src.engine_args import _resolve_cached_model_path, get_engine_args
+
+
+MODEL = "Qwen/Qwen3.6-27B-FP8"
+SNAPSHOT_HASH = "e89b16ebf1988b3d6befa7de50abc2d76f26eb09"
+
+
+def _make_cache(root, folder_name, snapshot=SNAPSHOT_HASH):
+    """Create a HF-style ``models--…/snapshots/<hash>/`` dir and return its path."""
+    snap_dir = os.path.join(root, folder_name, "snapshots", snapshot)
+    os.makedirs(snap_dir)
+    return snap_dir
+
+
+def _is_case_sensitive_fs(path):
+    """The lowercase-cache resolution only matters on case-sensitive filesystems.
+
+    On macOS (APFS, case-insensitive by default) ``models--Qwen--…`` and
+    ``models--qwen--…`` collide, so the resolver always sees the exact-case dir
+    as present. Production runs on Linux (case-sensitive), which is what these
+    tests exercise.
+    """
+    probe = os.path.join(path, "CaseProbe")
+    open(probe, "w").close()
+    try:
+        return not os.path.exists(os.path.join(path, "caseprobe"))
+    finally:
+        os.remove(probe)
+
+
+requires_case_sensitive_fs = pytest.mark.skipif(
+    not _is_case_sensitive_fs(os.environ.get("TMPDIR", "/tmp")),
+    reason="lowercase HF cache resolution only applies on case-sensitive filesystems",
+)
+
+
+@pytest.fixture
+def hf_cache(tmp_path, monkeypatch):
+    cache = tmp_path / "hub"
+    cache.mkdir()
+    monkeypatch.setenv("HUGGINGFACE_HUB_CACHE", str(cache))
+    # Make sure HF_HOME does not shadow the explicit cache dir during the test.
+    monkeypatch.delenv("HF_HOME", raising=False)
+    return cache
+
+
+class TestResolveCachedModelPath:
+    def test_exact_case_dir_returns_repo_id(self, hf_cache):
+        _make_cache(str(hf_cache), "models--Qwen--Qwen3.6-27B-FP8")
+        assert _resolve_cached_model_path(MODEL) == MODEL
+
+    def test_no_cache_returns_repo_id(self, hf_cache):
+        assert _resolve_cached_model_path(MODEL) == MODEL
+
+    def test_absolute_path_passthrough(self, hf_cache):
+        path = "/runpod-volume/some/local/model"
+        assert _resolve_cached_model_path(path) == path
+
+    @requires_case_sensitive_fs
+    def test_lowercase_dir_returns_snapshot_path(self, hf_cache):
+        snap = _make_cache(str(hf_cache), "models--qwen--qwen3.6-27b-fp8")
+        assert _resolve_cached_model_path(MODEL) == snap
+
+    def test_lowercase_dir_without_snapshots_returns_repo_id(self, hf_cache):
+        # Dir exists but has no snapshots subdir -> nothing to resolve to.
+        os.makedirs(os.path.join(str(hf_cache), "models--qwen--qwen3.6-27b-fp8"))
+        assert _resolve_cached_model_path(MODEL) == MODEL
+
+    @requires_case_sensitive_fs
+    def test_lowercase_dir_picks_latest_snapshot(self, hf_cache):
+        folder = "models--qwen--qwen3.6-27b-fp8"
+        _make_cache(str(hf_cache), folder, snapshot="aaaa")
+        latest = _make_cache(str(hf_cache), folder, snapshot="zzzz")
+        assert _resolve_cached_model_path(MODEL) == latest
+
+
+class TestGetEngineArgsServedName:
+    """Issue #310: served name must be decoupled from the resolved on-disk path."""
+
+    @pytest.fixture(autouse=True)
+    def base_env(self, monkeypatch):
+        # Avoid the network branch in _resolve_max_model_len.
+        monkeypatch.setenv("MAX_NUM_BATCHED_TOKENS", "2048")
+        monkeypatch.delenv("SERVED_MODEL_NAME", raising=False)
+        # Don't pick up a stray vLLM config file from the environment.
+        monkeypatch.setenv("VLLM_CONFIG_FILE", "/nonexistent-vllm-config.yaml")
+
+    @requires_case_sensitive_fs
+    def test_served_name_is_repo_id_when_path_rewritten(self, hf_cache, monkeypatch):
+        snap = _make_cache(str(hf_cache), "models--qwen--qwen3.6-27b-fp8")
+        monkeypatch.setenv("MODEL_NAME", MODEL)
+
+        result = get_engine_args()
+
+        assert result.model == snap  # weights load from the lowercase cache
+        assert result.served_model_name == MODEL  # API still serves the repo id
+
+    def test_served_name_untouched_when_no_rewrite(self, hf_cache, monkeypatch):
+        _make_cache(str(hf_cache), "models--Qwen--Qwen3.6-27B-FP8")
+        monkeypatch.setenv("MODEL_NAME", MODEL)
+
+        result = get_engine_args()
+
+        assert result.model == MODEL
+        assert result.served_model_name is None
+
+    def test_explicit_served_name_not_overridden(self, hf_cache, monkeypatch):
+        _make_cache(str(hf_cache), "models--qwen--qwen3.6-27b-fp8")
+        monkeypatch.setenv("MODEL_NAME", MODEL)
+        monkeypatch.setenv("SERVED_MODEL_NAME", "custom-name")
+
+        result = get_engine_args()
+
+        assert result.served_model_name == "custom-name"