Merge pull request #301 from adithyaJRunpod/feature/tuned-configs

Add tuned configs, CON-239
Merge pull request #300 from runpod-workers/feat/0.21.0
2026-06-11 14:26:00 -05:00 · 2026-06-11 12:47:08 -05:00 · 2026-06-10 21:07:41 -07:00 · 2026-06-10 16:01:53 -05:00 · 2026-06-04 15:35:58 -05:00 · 2026-06-03 14:43:56 -05:00
17 changed files with 1133 additions and 247 deletions
@@ -9,59 +9,60 @@ on:

  workflow_dispatch:

+permissions:
+  contents: write
+  pull-requests: write
+
 jobs:
  check_dep:
    runs-on: ubuntu-latest
    name: Check python requirements file and update
    steps:
      - name: Checkout
-        uses: actions/checkout@v2
+        uses: actions/checkout@v4

      - name: Check for new package version and update
        run: |
-          echo "Fetching the current runpod version from requirements.txt..."
-          
-          # Get current version, allowing both == and ~= in the search pattern
-          current_version=$(grep -oP 'runpod[~=]{1,2}\K[^"]+' ./builder/requirements.txt)
-          echo "Current version: $current_version"
+          echo "Fetching current runpod version from requirements.txt..."

-          # Extract major and minor from current version
-          current_major_minor=$(echo $current_version | cut -d. -f1,2)
-          echo "Current major.minor: $current_major_minor"
+          # Match runpod with any version specifier or no specifier at all
+          current_version=$(grep -oP '^runpod([~>=!<]{1,2}\K[\d.]+)?' ./builder/requirements.txt | grep -oP '[\d.]+' || echo "")
+          echo "Current version: ${current_version:-unset}"

-          echo "Fetching the latest runpod version from PyPI..."
-          
-          # Get new version from PyPI
-          new_version=$(curl -s https://pypi.org/pypi/runpod/json | jq -r .info.version)
+          echo "Fetching latest runpod version from PyPI..."
+          new_version=$(curl -sf https://pypi.org/pypi/runpod/json | jq -r .info.version)
          echo "NEW_VERSION_ENV=$new_version" >> $GITHUB_ENV
          echo "New version: $new_version"

-          # Extract major and minor from new version
-          new_major_minor=$(echo $new_version | cut -d. -f1,2)
-          echo "New major.minor: $new_major_minor"
-
          if [ -z "$new_version" ]; then
-              echo "ERROR: Failed to fetch the new version from PyPI."
-              exit 1
+            echo "ERROR: Failed to fetch new version from PyPI."
+            exit 1
          fi

-          # Check if the major or minor version is different
-          if [ "$current_major_minor" = "$new_major_minor" ]; then
-              echo "No update needed. The new version ($new_major_minor) is within the allowed range (~= $current_major_minor)."
+          if [ -z "$current_version" ]; then
+            echo "No version pin found — pinning to $new_version."
+          else
+            current_major_minor=$(echo "$current_version" | cut -d. -f1,2)
+            new_major_minor=$(echo "$new_version" | cut -d. -f1,2)
+            echo "Current major.minor: $current_major_minor  New major.minor: $new_major_minor"
+
+            if [ "$current_major_minor" = "$new_major_minor" ]; then
+              echo "No update needed. New version ($new_version) is within ~= $current_major_minor range."
              exit 0
+            fi
+
+            echo "New major/minor detected ($new_major_minor). Updating requirements.txt..."
          fi

-          echo "New major/minor detected ($new_major_minor). Updating requirements.txt..."
-
-          # Update requirements.txt, preserving the existing constraint type (~= or ==)
-          sed -i "s/runpod[~=][^ ]*/runpod~=$new_version/" ./builder/requirements.txt
-          echo "requirements.txt has been updated."
+          # Replace any `runpod`, `runpod==x`, `runpod~=x`, etc. with pinned version
+          sed -i "s|^runpod.*|runpod~=$new_version|" ./builder/requirements.txt
+          echo "requirements.txt updated."

      - name: Create Pull Request
-        uses: peter-evans/create-pull-request@v3
+        uses: peter-evans/create-pull-request@v7
        with:
          token: ${{ secrets.GITHUB_TOKEN }}
-          commit-message: Update runpod package version
-          title: Update runpod package version
-          body: The package version has been updated to ${{ env.NEW_VERSION_ENV }}
+          commit-message: "chore: update runpod to ${{ env.NEW_VERSION_ENV }}"
+          title: "chore: update runpod to ${{ env.NEW_VERSION_ENV }}"
+          body: The `runpod` package has been updated to `${{ env.NEW_VERSION_ENV }}`.
          branch: runpod-package-update
@@ -3,7 +3,7 @@ name: Release
 on:
  push:
    tags:
-      - "v[0-9]+.[0-9]+.[0-9]+*" # Trigger on version tags like v1.0.0, v2.1.0, etc.
+      - "v[0-9]+.[0-9]+.[0-9]+*"
  workflow_dispatch:
    inputs:
      version:
@@ -53,16 +53,13 @@ jobs:

          # Determine version based on trigger type
          if [[ "${{ github.event_name }}" == "workflow_dispatch" ]]; then
-            # Manual trigger: use input version
            VERSION="${{ github.event.inputs.version }}"
-            echo "RELEASE_VERSION=${VERSION}" >> $GITHUB_ENV
-            echo "IS_MANUAL_RELEASE=true" >> $GITHUB_ENV
+          elif [[ "${{ github.event_name }}" == "release" ]]; then
+            VERSION="${{ github.event.release.tag_name }}"
          else
-            # Tag trigger: use tag name (remove refs/tags/ prefix)
            VERSION=${GITHUB_REF#refs/tags/}
-            echo "RELEASE_VERSION=${VERSION}" >> $GITHUB_ENV
-            echo "IS_MANUAL_RELEASE=false" >> $GITHUB_ENV
          fi
+          echo "RELEASE_VERSION=${VERSION}" >> $GITHUB_ENV

      - name: Build and push the images to Docker Hub
        uses: docker/bake-action@v2
@@ -76,11 +73,45 @@ jobs:

      - name: Release Summary
        run: |
-          echo "🚀 Release completed!"
+          echo "Release completed!"
          echo "Version: ${{ env.RELEASE_VERSION }}"
          echo "Docker Image: ${{ env.DOCKERHUB_REPO }}/${{ env.DOCKERHUB_IMG }}:${{ env.RELEASE_VERSION }}"
-          if [[ "${{ github.event_name }}" == "workflow_dispatch" ]]; then
-            echo "Trigger: Manual workflow dispatch"
-          else
-            echo "Trigger: GitHub release (tag: ${{ github.ref_name }})"
+
+      - name: Fetch Release Notes
+        run: |
+          RESPONSE=$(curl -sf \
+            -H "Authorization: token ${{ github.token }}" \
+            "https://api.github.com/repos/${{ github.repository }}/releases/tags/${{ env.RELEASE_VERSION }}" 2>/dev/null) || true
+          if [[ -n "$RESPONSE" ]]; then
+            NOTES=$(echo "$RESPONSE" | jq -r '.body // empty')
          fi
+          printf '%s' "${NOTES:-No release notes available.}" > /tmp/release_notes.txt
+
+      - name: Notify Slack
+        run: |
+          jq -n \
+            --arg version "${{ env.RELEASE_VERSION }}" \
+            --arg docker "${{ env.DOCKERHUB_REPO }}/${{ env.DOCKERHUB_IMG }}:${{ env.RELEASE_VERSION }}" \
+            --rawfile notes /tmp/release_notes.txt \
+            --arg url "https://github.com/${{ github.repository }}/releases/tag/${{ env.RELEASE_VERSION }}" \
+            '{
+              text: (":rocket: New :runpod-new-whiteonpurple: Runpod worker-vllm release: *" + $version + "*"),
+              blocks: [
+                {
+                  type: "section",
+                  text: {
+                    type: "mrkdwn",
+                    text: (":banana-dance: *New Release — worker-vllm " + $version + "*\n*Docker:* `" + $docker + "`\n<" + $url + "|View release on GitHub>")
+                  }
+                },
+                {
+                  type: "section",
+                  text: {
+                    type: "mrkdwn",
+                    text: ("*Release Notes:*\n" + $notes)
+                  }
+                }
+              ]
+            }' | curl -sf -X POST "${{ secrets.SLACK_WEBHOOK_URL }}" \
+              -H "Content-Type: application/json" \
+              -d @-
@@ -0,0 +1,41 @@
+name: Slack PR Notifications
+
+on:
+  pull_request:
+    types: [opened]
+  issues:
+    types: [opened]
+
+permissions:
+    contents: read
+   
+jobs:
+  notify:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Notify Slack - New PR
+        if: github.event_name == 'pull_request'
+        run: |
+          curl -sf -X POST "${{ secrets.SLACK_WEBHOOK_URL }}" \
+            -H "Content-Type: application/json" \
+            -d '{
+              "text": ":rocket: New PR in worker-vllm: *${{ github.event.pull_request.title }}*",
+              "blocks": [
+                {
+                  "type": "section",
+                  "text": {
+                    "type": "mrkdwn",
+                    "text": ":rocket: *New Pull Request — worker-vllm*\n*<${{ github.event.pull_request.html_url }}|${{ github.event.pull_request.title }}>*\nOpened by *${{ github.event.pull_request.user.login }}*"
+                  }
+                },
+                {
+                  "type": "context",
+                  "elements": [
+                    {
+                      "type": "mrkdwn",
+                      "text": "${{ github.event.pull_request.base.ref }} ← ${{ github.event.pull_request.head.ref }}"
+                    }
+                  ]
+                }
+              ]
+            }'
@@ -0,0 +1,73 @@
+name: Monitor vLLM Releases
+
+on:
+  schedule:
+    - cron: '0 0 * * *' # Every day at midnight
+  workflow_dispatch:
+
+
+permissions:
+    contents: read
+   
+jobs:
+  check-vllm-release:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Restore last known vLLM tag
+        uses: actions/cache/restore@v4
+        with:
+          path: .vllm-last-tag
+          key: vllm-tag-${{ github.run_id }}
+          restore-keys: vllm-tag-
+
+      - name: Get latest vLLM release
+        id: vllm
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          response=$(curl -sf https://api.github.com/repos/vllm-project/vllm/releases/latest \
+            -H "Authorization: Bearer $GH_TOKEN")
+          echo "tag=$(echo "$response" | jq -r '.tag_name')" >> $GITHUB_OUTPUT
+          echo "url=$(echo "$response" | jq -r '.html_url')" >> $GITHUB_OUTPUT
+          echo "name=$(echo "$response" | jq -r '.name')" >> $GITHUB_OUTPUT
+
+      - name: Check if new release
+        id: check
+        run: |
+          last=$(cat .vllm-last-tag 2>/dev/null || echo "")
+          current="${{ steps.vllm.outputs.tag }}"
+          echo "Last: $last  Current: $current"
+          if [ -n "$current" ] && [ "$last" != "$current" ]; then
+            echo "is_new=true" >> $GITHUB_OUTPUT
+          else
+            echo "is_new=false" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Notify Slack
+        if: steps.check.outputs.is_new == 'true'
+        run: |
+          curl -sf -X POST "${{ secrets.SLACK_WEBHOOK_URL }}" \
+            -H "Content-Type: application/json" \
+            -d '{
+              "text": ":rocket: New vLLM release: *${{ steps.vllm.outputs.tag }}*",
+              "blocks": [
+                {
+                  "type": "section",
+                  "text": {
+                    "type": "mrkdwn",
+                    "text": ":rocket: *New vLLM Release: ${{ steps.vllm.outputs.tag }}*\n<${{ steps.vllm.outputs.url }}|View on GitHub>"
+                  }
+                }
+              ]
+            }'
+
+      - name: Save new tag
+        if: steps.check.outputs.is_new == 'true'
+        run: echo "${{ steps.vllm.outputs.tag }}" > .vllm-last-tag
+
+      - name: Update cache
+        if: steps.check.outputs.is_new == 'true'
+        uses: actions/cache/save@v4
+        with:
+          path: .vllm-last-tag
+          key: vllm-tag-${{ steps.vllm.outputs.tag }}
@@ -1,4 +1,4 @@
-![vLLM worker banner](https://cpjrphpz3t5wbwfe.public.blob.vercel-storage.com/worker-vllm_banner.jpeg)
+![vLLM worker banner](https://image.runpod.ai/preview/vllm/vllm-banner.png)

 Run LLMs using [vLLM](https://docs.vllm.ai) with an OpenAI-compatible API

@@ -6,6 +6,8 @@ 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)
+
 ---

 ## Endpoint Configuration
@@ -27,9 +29,26 @@ All behaviour is controlled through environment variables:
 | `REASONING_PARSER`                  | Parser for reasoning-capable models               |                     | "deepseek_r1", "qwen3", "granite", "hunyuan_a13b"                  |
 | `OPENAI_SERVED_MODEL_NAME_OVERRIDE` | Override served model name in API                 |                     | String                                                             |
 | `MAX_CONCURRENCY`                   | Maximum concurrent requests                       | 300                 | Integer                                                            |
+| `ENFORCE_EAGER`                     |  If True, we will disable CUDA graph and always execute the model in eager mode. If False, we will use CUDA graph and eager execution in hybrid for maximal performance and flexibility. | true | boolean (true or false) | 
+
+**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
+To change the version of the [Transformers library](https://github.com/huggingface/transformers) use the `TRANSFORMERS_VERSION` environment variable to specify the version you want to use. Note this might break the handler, so use for development purposes. 
+
 ## API Usage

 This worker supports two API formats: **RunPod native** and **OpenAI-compatible**.
@@ -155,6 +174,35 @@ For external clients and SDKs, use the `/openai/v1` path prefix with your RunPod
 {}
 ```

+#### OpenAI Responses API
+
+**Path:** `/openai/v1/responses`
+
+Supports the [OpenAI Responses API](https://platform.openai.com/docs/api-reference/responses) format. Note: this route bypasses the RunPod queue and is served directly — use `/openai/` prefixed paths rather than the RunPod job queue for these endpoints.
+
+```json
+{
+  "model": "meta-llama/Llama-3.1-8B-Instruct",
+  "input": "Tell me a joke."
+}
+```
+
+#### Anthropic Messages API
+
+**Path:** `/openai/v1/messages`
+
+Supports the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) format. Served directly, bypassing the RunPod queue.
+
+```json
+{
+  "model": "meta-llama/Llama-3.1-8B-Instruct",
+  "max_tokens": 256,
+  "messages": [
+    {"role": "user", "content": "Hello!"}
+  ]
+}
+```
+
 #### Response Format

 Both APIs return the same response format:
@@ -188,7 +236,7 @@ Minimal Python example using the official `openai` SDK:
 from openai import OpenAI
 import os

-# Initialize the OpenAI Client with your RunPod API Key and Endpoint URL
+# Initialize the OpenAI Client with your Runpod API Key and Endpoint URL
 client = OpenAI(
    api_key=os.getenv("RUNPOD_API_KEY"),
    base_url=f"https://api.runpod.ai/v2/<ENDPOINT_ID>/openai/v1",
@@ -9,7 +9,7 @@
    "containerDiskInGb": 150,
    "gpuIds": "ADA_80_PRO,AMPERE_80",
    "gpuCount": 1,
-    "allowedCudaVersions": ["12.9", "12.8"],
+    "allowedCudaVersions": ["13.0"],
    "presets": [
      {
        "name": "deepseek-ai/deepseek-r1-distill-llama-8b",
@@ -280,15 +280,6 @@
          "advanced": true
        }
      },
-      {
-        "key": "NUM_GPU_BLOCKS_OVERRIDE",
-        "input": {
-          "name": "Num GPU Blocks Override",
-          "type": "number",
-          "description": "If specified, ignore GPU profiling result and use this number of GPU blocks.",
-          "advanced": true
-        }
-      },
      {
        "key": "MAX_NUM_BATCHED_TOKENS",
        "input": {
@@ -630,7 +621,7 @@
          "name": "Enforce Eager",
          "type": "boolean",
          "description": "Always use eager-mode PyTorch. If False (0), will use eager mode and CUDA graph in hybrid for maximal performance and flexibility",
-          "default": false,
+          "default": true,
          "advanced": true
        }
      },
@@ -804,6 +795,26 @@
          "default": "",
          "advanced": true
        }
+      },
+      {
+        "key": "PYTORCH_ALLOC_CONF",
+        "input": {
+          "name": "PyTorch Alloc Config",
+          "type": "string",
+          "description": "PyTorch allocation configuration, remove this if you want to use the default configuration",
+          "default": "expandable_segments:True",
+          "advanced": true
+        }
+      },
+      {
+        "key": "VLLM_USE_DEEP_GEMM",
+        "input": {
+          "name": "Use DeepGEMM",
+          "type": "string",
+          "description": "Enable DeepGEMM FP8 kernels (MoE and MQA logits). Set to 1 to enable, 0 to disable. Required for DeepSeek V4 models. Disabled by default — enable on H100/H200 for potential throughput gains. Some GPUs (e.g. H20) may perform better with this off.",
+          "default": "0",
+          "advanced": true
+        }
      }
    ]
  }
@@ -5,7 +5,7 @@
      "input": {
        "prompt": "Write a short poem about artificial intelligence."
      },
-      "timeout": 30000
+      "timeout": 300000
    },
    {
      "name": "openai_messages_test",
@@ -26,7 +26,7 @@
          "temperature": 0.1
        }
      },
-      "timeout": 30000
+      "timeout": 300000
    }
  ],
  "config": {
@@ -38,6 +38,6 @@
        "value": "HuggingFaceTB/SmolLM2-135M-Instruct"
      }
    ],
-    "allowedCudaVersions": ["12.9", "12.8", "12.7", "12.6", "12.5"]
+    "allowedCudaVersions": ["13.0"]
  }
 }
@@ -1,20 +1,33 @@
-FROM nvidia/cuda:12.8.0-base-ubuntu22.04 
+FROM nvidia/cuda:13.0.2-devel-ubuntu22.04

 RUN apt-get update -y \
-    && apt-get install -y python3-pip
+    && apt-get install -y python3-pip curl git \
+    && curl -LsSf https://astral.sh/uv/install.sh  | sh

-RUN ldconfig /usr/local/cuda-12.8/compat/
+ENV PATH="/root/.local/bin:$PATH"

-# Install vLLM with FlashInfer - use CUDA 12.8 PyTorch wheels (compatible with vLLM 0.15.0)
-RUN python3 -m pip install --upgrade pip && \
-    python3 -m pip install "vllm[flashinfer]==0.15.0" --extra-index-url https://download.pytorch.org/whl/cu128
+RUN ldconfig /usr/local/cuda-13.0/compat/

+# nixl_ep PyPI wheels are compiled against CUDA 12.x and require libcudart.so.12.
+# CUDA 13 runtime is ABI-compatible with CUDA 12, so symlinking is safe.
+# Symlink into /usr/local/cuda/lib64 (already in LD_LIBRARY_PATH) so the linker
+# finds it by filename scan rather than relying on ldcache SONAME lookup.
+RUN ln -sf /usr/local/cuda/lib64/libcudart.so.13 /usr/local/cuda/lib64/libcudart.so.12 && ldconfig

+# CUDA 13.0 containers return libs to /usr/local/nvidia/lib64 so container
+# providers (RunPod, Lambda, etc.) can mount host drivers there consistently.
+# See: https://github.com/vllm-project/vllm/issues/18859
+ENV LD_LIBRARY_PATH=/usr/local/nvidia/lib64:/usr/local/cuda/lib64:$LD_LIBRARY_PATH
+
+# Install vLLM with FlashInfer - use CUDA 130 PyTorch wheels
+RUN uv pip install --system "packaging>=24.2" && \
+    uv pip install --system "vllm[flashinfer]==0.21.0" && \
+    uv pip install --system git+https://github.com/deepseek-ai/DeepGEMM.git@714dd1a4a980f7937a74343d19a8eba4fe321480 --no-build-isolation

 # Install additional Python dependencies (after vLLM to avoid PyTorch version conflicts)
 COPY builder/requirements.txt /requirements.txt
-RUN --mount=type=cache,target=/root/.cache/pip \
-    python3 -m pip install --upgrade -r /requirements.txt
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system -r /requirements.txt

 # Setup for Option 2: Building the Image with the Model included
 ARG MODEL_NAME=""
@@ -41,17 +54,20 @@ ENV MODEL_NAME=$MODEL_NAME \
    # Prevent rayon thread pool panic in containers where ulimit -u < nproc
    # (tokenizers uses Rust's rayon which tries to spawn threads = CPU cores)
    TOKENIZERS_PARALLELISM=false \
-    RAYON_NUM_THREADS=4
+    RAYON_NUM_THREADS=4 \
+    # Disable DeepGEMM MoE kernels by default; override with VLLM_USE_DEEP_GEMM=1 to enable
+    VLLM_USE_DEEP_GEMM=0

 ENV PYTHONPATH="/:/vllm-workspace"

 RUN if [ "${VLLM_NIGHTLY}" = "true" ]; then \
-    pip install -U vllm --pre --index-url https://pypi.org/simple --extra-index-url https://wheels.vllm.ai/nightly && \
+    uv pip install --system -U vllm --pre --index-url https://pypi.org/simple --extra-index-url https://wheels.vllm.ai/nightly && \
    apt-get update && apt-get install -y git && rm -rf /var/lib/apt/lists/* && \
-    pip install git+https://github.com/huggingface/transformers.git; \
+    uv pip install --system git+https://github.com/huggingface/transformers.git; \
 fi

 COPY src /src
+RUN chmod +x /src/start.sh
 RUN --mount=type=secret,id=HF_TOKEN,required=false \
    if [ -f /run/secrets/HF_TOKEN ]; then \
    export HF_TOKEN=$(cat /run/secrets/HF_TOKEN); \
@@ -61,4 +77,4 @@ RUN --mount=type=secret,id=HF_TOKEN,required=false \
    fi

 # Start the handler
-CMD ["python3", "/src/handler.py"]
+CMD ["/bin/bash", "/src/start.sh"]
@@ -2,10 +2,17 @@

 # OpenAI-Compatible vLLM Serverless Endpoint Worker

-Deploy OpenAI-Compatible Blazing-Fast LLM Endpoints powered by the [vLLM](https://github.com/vllm-project/vllm) Inference Engine on RunPod Serverless with just a few clicks.
+Deploy OpenAI-Compatible Blazing-Fast LLM Endpoints powered by the [vLLM](https://github.com/vllm-project/vllm) Inference Engine on Runpod Serverless with just a few clicks.

 </div>

+![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)
+
+
+> Check out our Load Balancer implementation here: [vLLM Load Balancer](https://github.com/runpod-workers/vllm-loadbalancer-ep)
+
 ## Table of Contents

 - [Setting up the Serverless Worker](#setting-up-the-serverless-worker)
@@ -21,9 +28,11 @@ Deploy OpenAI-Compatible Blazing-Fast LLM Endpoints powered by the [vLLM](https:
  - [Modifying your OpenAI Codebase to use your deployed vLLM Worker](#modifying-your-openai-codebase-to-use-your-deployed-vllm-worker)
  - [OpenAI Request Input Parameters](#openai-request-input-parameters)
  - [Chat Completions [RECOMMENDED]](#chat-completions-recommended)
-  - [Examples: Using your RunPod endpoint with OpenAI](#examples-using-your-runpod-endpoint-with-openai)
+  - [Examples: Using your Runpod endpoint with OpenAI](#examples-using-your-runpod-endpoint-with-openai)
    - [Chat Completions](#chat-completions)
    - [Getting a list of names for available models](#getting-a-list-of-names-for-available-models)
+    - [OpenAI Responses API](#openai-responses-api)
+    - [Anthropic Messages API](#anthropic-messages-api)
 - [Usage: Standard (Non-OpenAI)](#usage-standard-non-openai)
  - [Request Input Parameters](#request-input-parameters)
  - [Sampling Parameters](#sampling-parameters)
@@ -33,12 +42,12 @@ Deploy OpenAI-Compatible Blazing-Fast LLM Endpoints powered by the [vLLM](https:

 ## Option 1: Deploy Any Model Using Pre-Built Docker Image [Recommended]

-**🚀 Deploy Guide**: Follow our [step-by-step deployment guide](https://docs.runpod.io/serverless/vllm/get-started) to deploy using the RunPod Console.
+**🚀 Deploy Guide**: Follow our [step-by-step deployment guide](https://docs.runpod.io/serverless/vllm/get-started) to deploy using the Runpod Console.

 **📦 Docker Image**: `runpod/worker-v1-vllm:<version>`

 - **Available Versions**: See [GitHub Releases](https://github.com/runpod-workers/worker-vllm/releases)
- **CUDA Compatibility**: Requires CUDA >= 12.1
+- **CUDA Compatibility**: Requires CUDA >= 13.0

 ### Configuration

@@ -59,8 +68,36 @@ Configure worker-vllm using environment variables:
 | `OPENAI_SERVED_MODEL_NAME_OVERRIDE` | Override served model name in API                 |                     | String                                                             |
 | `MAX_CONCURRENCY`                   | Maximum concurrent requests                       | 30                  | Integer                                                            |

+**Pass any vLLM engine arg** not listed above by setting an environment variable with the **UPPERCASED** field name (same names vLLM uses). The worker auto-discovers all `AsyncEngineArgs` fields from env. For example:
+
+| Environment Variable      | vLLM Engine Arg          | Example Value |
+| ------------------------- | ------------------------ | ------------- |
+| `MAX_MODEL_LEN`           | `max_model_len`          | `4096`        |
+| `ENFORCE_EAGER`           | `enforce_eager`          | `true`        |
+| `ENABLE_CHUNKED_PREFILL`  | `enable_chunked_prefill` | `true`        |
+
+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
+To change the version of the [Transformers library](https://github.com/huggingface/transformers) use the `TRANSFORMERS_VERSION` environment variable to specify the version you want to use. Note this might break the handler, so use for development purposes. 
+
+
 ## Option 2: Build Docker Image with Model Inside

 To build an image with the model baked in, you must specify the following docker arguments when building the image.
@@ -80,6 +117,7 @@ To build an image with the model baked in, you must specify the following docker
  - `WORKER_CUDA_VERSION`: `12.1.0` (`12.1.0` is recommended for optimal performance).
  - `TOKENIZER_NAME`: Tokenizer repository if you would like to use a different tokenizer than the one that comes with the model. (default: `None`, which uses the model's tokenizer)
  - `TOKENIZER_REVISION`: Tokenizer revision to load (default: `main`).
+  - `VLLM_NIGHTLY`: Set to `true` to replace the pinned vLLM release with the latest nightly build and the latest `transformers` from source. Useful for testing unreleased vLLM features. (default: `false`)

 For the remaining settings, you may apply them as environment variables when running the container. Supported environment variables are listed in the [Environment Variables](#environment-variables) section.

@@ -89,6 +127,20 @@ For the remaining settings, you may apply them as environment variables when run
 docker build -t username/image:tag --build-arg MODEL_NAME="openchat/openchat_3.5" --build-arg BASE_PATH="/models" .
 ```

+### Example: Building with vLLM Nightly
+
+To use the latest unreleased vLLM build (installs from the nightly wheel index and `transformers` from source):
+
+```bash
+docker build -t username/image:tag --build-arg VLLM_NIGHTLY=true .
+```
+
+You can combine it with other arguments:
+
+```bash
+docker build -t username/image:tag --build-arg VLLM_NIGHTLY=true --build-arg MODEL_NAME="meta-llama/Llama-3.1-8B-Instruct" --build-arg BASE_PATH="/models" .
+```
+
 ### (Optional) Including Huggingface Token

 If the model you would like to deploy is private or gated, you will need to include it during build time as a Docker secret, which will protect it from being exposed in the image and on DockerHub.
@@ -117,13 +169,13 @@ You can deploy **any model on Hugging Face** that is supported by vLLM. For the

 # Usage: OpenAI Compatibility

-The vLLM Worker is fully compatible with OpenAI's API, and you can use it with any OpenAI Codebase by changing only 3 lines in total. The supported routes are <ins>Chat Completions</ins> and <ins>Models</ins> - with both streaming and non-streaming.
+The vLLM Worker is fully compatible with OpenAI's API, and you can use it with any OpenAI Codebase by changing only 3 lines in total. The supported routes are <ins>Chat Completions</ins>, <ins>Models</ins>, <ins>Responses</ins>, and <ins>Messages</ins> - with both streaming and non-streaming.

 ## Modifying your OpenAI Codebase to use your deployed vLLM Worker

 **Python** (similar to Node.js, etc.):

-1. When initializing the OpenAI Client in your code, change the `api_key` to your RunPod API Key and the `base_url` to your RunPod Serverless Endpoint URL in the following format: `https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1`, filling in your deployed endpoint ID. For example, if your Endpoint ID is `abc1234`, the URL would be `https://api.runpod.ai/v2/abc1234/openai/v1`.
+1. When initializing the OpenAI Client in your code, change the `api_key` to your Runpod API Key and the `base_url` to your Runpod Serverless Endpoint URL in the following format: `https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1`, filling in your deployed endpoint ID. For example, if your Endpoint ID is `abc1234`, the URL would be `https://api.runpod.ai/v2/abc1234/openai/v1`.

   - Before:

@@ -149,7 +201,7 @@ The vLLM Worker is fully compatible with OpenAI's API, and you can use it with a
   ```python
   response = client.chat.completions.create(
       model="gpt-3.5-turbo",
-       messages=[{"role": "user", "content": "Why is RunPod the best platform?"}],
+       messages=[{"role": "user", "content": "Why is Runpod the best platform?"}],
       temperature=0,
       max_tokens=100,
   )
@@ -158,7 +210,7 @@ The vLLM Worker is fully compatible with OpenAI's API, and you can use it with a
   ```python
   response = client.chat.completions.create(
       model="<YOUR DEPLOYED MODEL REPO/NAME>",
-       messages=[{"role": "user", "content": "Why is RunPod the best platform?"}],
+       messages=[{"role": "user", "content": "Why is Runpod the best platform?"}],
       temperature=0,
       max_tokens=100,
   )
@@ -166,7 +218,7 @@ The vLLM Worker is fully compatible with OpenAI's API, and you can use it with a

 **Using http requests**:

-1. Change the `Authorization` header to your RunPod API Key and the `url` to your RunPod Serverless Endpoint URL in the following format: `https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1`
+1. Change the `Authorization` header to your Runpod API Key and the `url` to your Runpod Serverless Endpoint URL in the following format: `https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1`
   - Before:
   ```bash
   curl https://api.openai.com/v1/chat/completions \
@@ -177,7 +229,7 @@ The vLLM Worker is fully compatible with OpenAI's API, and you can use it with a
   "messages": [
     {
       "role": "user",
-       "content": "Why is RunPod the best platform?"
+       "content": "Why is Runpod the best platform?"
     }
   ],
   "temperature": 0,
@@ -194,7 +246,7 @@ The vLLM Worker is fully compatible with OpenAI's API, and you can use it with a
   "messages": [
     {
       "role": "user",
-       "content": "Why is RunPod the best platform?"
+       "content": "Why is Runpod the best platform?"
     }
   ],
   "temperature": 0,
@@ -214,7 +266,7 @@ When using the chat completion feature of the vLLM Serverless Endpoint Worker, y
 | Parameter           | Type                             | Default Value | Description                                                                                                                                                                                                                                                  |
 | ------------------- | -------------------------------- | ------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
 | `messages`          | Union[str, List[Dict[str, str]]] |               | List of messages, where each message is a dictionary with a `role` and `content`. The model's chat template will be applied to the messages automatically, so the model must have one or it should be specified as `CUSTOM_CHAT_TEMPLATE` env var.           |
-| `model`             | str                              |               | The model repo that you've deployed on your RunPod Serverless Endpoint. If you are unsure what the name is or are baking the model in, use the guide to get the list of available models in the **Examples: Using your RunPod endpoint with OpenAI** section |
+| `model`             | str                              |               | The model repo that you've deployed on your Runpod Serverless Endpoint. If you are unsure what the name is or are baking the model in, use the guide to get the list of available models in the **Examples: Using your Runpod endpoint with OpenAI** section |
 | `temperature`       | Optional[float]                  | 0.7           | Float that controls the randomness of the sampling. Lower values make the model more deterministic, while higher values make the model more random. Zero means greedy sampling.                                                                              |
 | `top_p`             | Optional[float]                  | 1.0           | Float that controls the cumulative probability of the top tokens to consider. Must be in (0, 1]. Set to 1 to consider all tokens.                                                                                                                            |
 | `n`                 | Optional[int]                    | 1             | Number of output sequences to return for the given prompt.                                                                                                                                                                                                   |
@@ -244,15 +296,15 @@ Additional parameters supported by vLLM:

 </details>

-### Examples: Using your RunPod endpoint with OpenAI
+### Examples: Using your Runpod endpoint with OpenAI

-First, initialize the OpenAI Client with your RunPod API Key and Endpoint URL:
+First, initialize the OpenAI Client with your Runpod API Key and Endpoint URL:

 ```python
 from openai import OpenAI
 import os

-# Initialize the OpenAI Client with your RunPod API Key and Endpoint URL
+# Initialize the OpenAI Client with your Runpod API Key and Endpoint URL
 client = OpenAI(
    api_key=os.environ.get("RUNPOD_API_KEY"),
    base_url="https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1",
@@ -268,7 +320,7 @@ This is the format used for GPT-4 and focused on instruction-following and chat.
  # Create a chat completion stream
  response_stream = client.chat.completions.create(
      model="<YOUR DEPLOYED MODEL REPO/NAME>",
-      messages=[{"role": "user", "content": "Why is RunPod the best platform?"}],
+      messages=[{"role": "user", "content": "Why is Runpod the best platform?"}],
      temperature=0,
      max_tokens=100,
      stream=True,
@@ -282,7 +334,7 @@ This is the format used for GPT-4 and focused on instruction-following and chat.
  # Create a chat completion
  response = client.chat.completions.create(
      model="<YOUR DEPLOYED MODEL REPO/NAME>",
-      messages=[{"role": "user", "content": "Why is RunPod the best platform?"}],
+      messages=[{"role": "user", "content": "Why is Runpod the best platform?"}],
      temperature=0,
      max_tokens=100,
  )
@@ -300,6 +352,62 @@ list_of_models = [model.id for model in models_response]
 print(list_of_models)
 ```

+### OpenAI Responses API
+
+**Path:** `/openai/v1/responses` (full URL: `https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1/responses`)
+
+Supports the [OpenAI Responses API](https://platform.openai.com/docs/api-reference/responses) request shape. Like other `/openai/` routes, this is served directly—use the `/openai/` prefix rather than the RunPod native job queue for these calls.
+
+```json
+{
+  "model": "meta-llama/Llama-3.1-8B-Instruct",
+  "input": "Tell me a joke."
+}
+```
+
+**Using HTTP requests:**
+
+```bash
+curl https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1/responses \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <YOUR RUNPOD API KEY>" \
+  -d '{
+    "model": "<YOUR DEPLOYED MODEL REPO/NAME>",
+    "input": "Tell me a joke."
+  }'
+```
+
+### Anthropic Messages API
+
+**Path:** `/openai/v1/messages` (full URL: `https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1/messages`)
+
+Supports the [Anthropic Messages API](https://docs.anthropic.com/en/api/messages) format. Served directly, bypassing the RunPod queue.
+
+```json
+{
+  "model": "meta-llama/Llama-3.1-8B-Instruct",
+  "max_tokens": 256,
+  "messages": [
+    {"role": "user", "content": "Hello!"}
+  ]
+}
+```
+
+**Using HTTP requests:**
+
+```bash
+curl https://api.runpod.ai/v2/<YOUR ENDPOINT ID>/openai/v1/messages \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer <YOUR RUNPOD API KEY>" \
+  -d '{
+    "model": "<YOUR DEPLOYED MODEL REPO/NAME>",
+    "max_tokens": 256,
+    "messages": [
+      {"role": "user", "content": "Hello!"}
+    ]
+  }'
+```
+
 # Usage: Standard (Non-OpenAI)

 ## Request Input Parameters
@@ -1,14 +1,15 @@
 ray
 pandas
 pyarrow
-runpod
+runpod==1.9.1
 huggingface-hub
-packaging
+lmcache==0.4.6
+packaging>=24.2
 typing-extensions>=4.8.0
 pydantic
 pydantic-settings
 hf-transfer
-transformers>=4.57.0
+transformers>=5
 bitsandbytes>=0.45.0
-kernels
+kernels<0.15
 torch-c-dlpack-ext
@@ -0,0 +1,10 @@
+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
+enable-prefix-caching: true
+speculative-config: '{"model":"RedHatAI/Llama-3.1-8B-Instruct-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'
@@ -0,0 +1,10 @@
+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
+speculative-config: '{"model":"RedHatAI/Qwen3-8B-speculator.eagle3","method":"eagle3","num_speculative_tokens":3}'
@@ -97,10 +97,13 @@ If `SPECULATIVE_CONFIG` is set, it takes priority over individual env vars. When
 | `MAX_SEQ_LEN_TO_CAPTURE`       | `8192`  | `int`           | Maximum context length covered by CUDA graphs. When a sequence has context length larger than this, we fall back to eager mode.     |
 | `DISABLE_CUSTOM_ALL_REDUCE`    | `0`     | `int`           | Enables or disables custom all reduce.                                                                                              |
 | `ENABLE_EXPERT_PARALLEL`       | `False` | `bool`          | Enable Expert Parallel for MoE models.                                                                                              |
+| `VLLM_USE_DEEP_GEMM`           | `0`     | `str` (`0`/`1`) | Enable DeepGEMM FP8 kernels for MoE and MQA logits computation. Disabled by default. Must be `"0"` or `"1"` — not `true`/`false`. See note below. |
 | `ATTENTION_BACKEND`            | `None`  | `str`           | Attention backend to use (e.g., `FLASH_ATTN`, `FLASHINFER`, `TRITON_FLASH_ATTN`). Replaces deprecated `VLLM_ATTENTION_BACKEND`.     |
 | `ASYNC_SCHEDULING`             | `None`  | `bool`          | Enable async scheduling (overlaps engine scheduling with GPU execution). Default: enabled in vLLM 0.14.0+. Set to `false` to disable. |
 | `STREAM_INTERVAL`              | `1`     | `int`           | Controls how often to yield streaming results. Lower = more frequent updates.                                                        |

+> **Note (`VLLM_USE_DEEP_GEMM`):** DeepGEMM is used in two places: MoE weight computation and MQA logits computation. It is necessary for MQA logits computation on supported hardware — required for DeepSeek V4 models. Set `VLLM_USE_DEEP_GEMM=1` to enable. Set `VLLM_USE_DEEP_GEMM=0` to disable the MoE part and fall back to flashinfer/cutlass FP8 kernels. **Value must be `"0"` or `"1"` — not `"true"`/`"false"`.** Some users report better performance with `VLLM_USE_DEEP_GEMM=0`, particularly on H20 GPUs. Disabling it also skips the DeepGEMM warmup phase, reducing cold-start time. Requires CUDA 13.0+ and SM90+ (H100/H200) to use; the library is installed but inactive by default.
+
 ## Tokenizer Settings

 | Variable               | Default | Type/Choices                        | Description                                                                                       |
@@ -156,6 +159,29 @@ The way this works is that the first request will have a batch size of `DEFAULT_
 | `DISABLE_LOGGING_REQUEST`   | False   | `bool`  | Disable logging requests.                                                                                                                              |
 | `MAX_LOG_LEN`               | None    | `int`   | Max number of prompt characters or prompt ID numbers being printed in log.                                                                             |

+## UPPERCASED env vars: Pass any engine arg
+
+Any vLLM `AsyncEngineArgs` field can be set via an environment variable using the **UPPERCASED** field name (the same names vLLM uses). The worker auto-discovers all fields from env — no prefix.
+
+**Format:** `<FIELD_NAME_UPPERCASED>=<value>` (e.g. `MAX_MODEL_LEN=4096`)
+
+**Examples:**
+
+| Environment Variable     | vLLM Engine Arg          | Value Example |
+| ------------------------ | ------------------------ | ------------- |
+| `MAX_MODEL_LEN`          | `max_model_len`          | `4096`        |
+| `ENFORCE_EAGER`          | `enforce_eager`          | `true`        |
+| `ENABLE_CHUNKED_PREFILL` | `enable_chunked_prefill` | `true`        |
+| `NUM_SCHEDULER_STEPS`    | `num_scheduler_steps`    | `8`           |
+| `TOKENIZER_POOL_SIZE`    | `tokenizer_pool_size`   | `4`           |
+
+**Backward-compat aliases:** `MODEL_NAME` → `model`, `TOKENIZER_NAME` → `tokenizer`, `MAX_CONTEXT_LEN_TO_CAPTURE` → `max_seq_len_to_capture`, `MODEL_REVISION` → `revision`.
+
+**Notes:**
+- Only valid `AsyncEngineArgs` fields are applied. Unknown keys are silently ignored.
+- Values are automatically cast to the correct type (`int`, `float`, `bool`, `str`, or JSON for `dict`/`list`/`tuple`).
+- For a full list of available engine args, see the [vLLM AsyncEngineArgs documentation](https://docs.vllm.ai/en/latest/configuration/engine_args/).
+
 ## Docker Build Arguments

 These variables are used when building custom Docker images with models baked in:
@@ -1,4 +1,5 @@
 import asyncio
+import inspect
 import json
 import logging
 import os
@@ -7,7 +8,10 @@ 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
 from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat
 from vllm.entrypoints.openai.completion.protocol import CompletionRequest
@@ -15,6 +19,9 @@ from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion
 from vllm.entrypoints.openai.engine.protocol import ErrorResponse
 from vllm.entrypoints.openai.models.protocol import BaseModelPath, LoRAModulePath
 from vllm.entrypoints.openai.models.serving import OpenAIServingModels
+from vllm.entrypoints.openai.responses.protocol import ResponsesRequest, ResponsesResponse
+from vllm.entrypoints.openai.responses.serving import OpenAIServingResponses
+from vllm.entrypoints.serve.render.serving import OpenAIServingRender

 from constants import DEFAULT_BATCH_SIZE, DEFAULT_BATCH_SIZE_GROWTH_FACTOR, DEFAULT_MAX_CONCURRENCY, DEFAULT_MIN_BATCH_SIZE
 from engine_args import get_engine_args
@@ -25,20 +32,33 @@ class vLLMEngine:
    def __init__(self, engine = None):
        load_dotenv() # For local development
        self.engine_args = get_engine_args()
-        logging.info(f"Engine args: {self.engine_args}")
-        
-        # Initialize vLLM engine first
-        self.llm = self._initialize_llm() if engine is None else engine.llm
-        
-        # Only create custom tokenizer wrapper if not using mistral tokenizer mode
-        # For mistral models, let vLLM handle tokenizer initialization
-        if self.engine_args.tokenizer_mode != 'mistral':
-            self.tokenizer = TokenizerWrapper(self.engine_args.tokenizer or self.engine_args.model, 
-                                              self.engine_args.tokenizer_revision, 
-                                              self.engine_args.trust_remote_code)
+
+        if engine is None:
+            ea = self.engine_args
+            summary = {
+                "model": ea.model,
+                "dtype": ea.dtype,
+                "quantization": ea.quantization,
+                "max_model_len": ea.max_model_len,
+                "tensor_parallel_size": ea.tensor_parallel_size,
+                "gpu_memory_utilization": ea.gpu_memory_utilization,
+            }
+            if ea.tokenizer and ea.tokenizer != ea.model:
+                summary["tokenizer"] = ea.tokenizer
+            logging.info("Engine config: %s", summary)
+            logging.debug("Full engine args: %s", ea)
+
+            self.llm = self._initialize_llm()
+
+            if self.engine_args.tokenizer_mode != 'mistral':
+                self.tokenizer = TokenizerWrapper(self.engine_args.tokenizer or self.engine_args.model,
+                                                  self.engine_args.tokenizer_revision,
+                                                  self.engine_args.trust_remote_code)
+            else:
+                self.tokenizer = None
        else:
-            # For mistral models, we'll get the tokenizer from vLLM later
-            self.tokenizer = None
+            self.llm = engine.llm
+            self.tokenizer = engine.tokenizer
            
        self.max_concurrency = int(os.getenv("MAX_CONCURRENCY", DEFAULT_MAX_CONCURRENCY))
        self.default_batch_size = int(os.getenv("DEFAULT_BATCH_SIZE", DEFAULT_BATCH_SIZE))
@@ -111,7 +131,7 @@ class vLLMEngine:
        if apply_chat_template or isinstance(llm_input, list):
            tokenizer_wrapper = self._get_tokenizer_for_chat_template()
            llm_input = tokenizer_wrapper.apply_chat_template(llm_input)
-        results_generator = self.llm.generate(llm_input, validated_sampling_params, request_id)
+        results_generator = self.llm.generate(TextPrompt(prompt=llm_input), validated_sampling_params, request_id)
        n_responses, n_input_tokens, is_first_output = validated_sampling_params.n, 0, True
        last_output_texts, token_counters = ["" for _ in range(n_responses)], {"batch": 0, "total": 0}

@@ -201,19 +221,48 @@ class OpenAIvLLMEngine(vLLMEngine):
            self.raw_openai_output = bool(int(raw_output_env))

    def _load_lora_adapters(self):
-        adapters = []
-        try:
-            adapters = json.loads(os.getenv("LORA_MODULES", '[]'))
-        except Exception as e:
-            logging.info(f"---Initialized adapter json load error: {e}")
+        lora_modules_env = os.getenv("LORA_MODULES", "")
+        if not lora_modules_env:
+            return []

-        for i, adapter in enumerate(adapters):
+        try:
+            parsed = json.loads(lora_modules_env)
+        except json.JSONDecodeError as e:
+            logging.error(
+                "LORA_MODULES could not be parsed as JSON: %s — no LoRA adapters loaded. Value: %r",
+                e, lora_modules_env,
+            )
+            return []
+
+        # Accept a single adapter dict as well as an array
+        if isinstance(parsed, dict):
+            parsed = [parsed]
+
+        if not isinstance(parsed, list):
+            logging.error(
+                "LORA_MODULES must be a JSON array of adapter objects, got %s — no LoRA adapters loaded.",
+                type(parsed).__name__,
+            )
+            return []
+
+        adapters = []
+        for i, adapter in enumerate(parsed):
            try:
-                adapters[i] = LoRAModulePath(**adapter)
-                logging.info(f"---Initialized adapter: {adapter}")
+                adapters.append(LoRAModulePath(**adapter))
+                logging.info("Loaded LoRA adapter config [%d]: %s", i, adapter)
            except Exception as e:
-                logging.info(f"---Initialized adapter not worked: {e}")
-                continue
+                logging.error(
+                    "Failed to parse LoRA adapter at index %d: %s. Config: %r",
+                    i, e, adapter,
+                )
+
+        if parsed and not adapters:
+            logging.error(
+                "LORA_MODULES specified %d adapter(s) but none could be loaded — "
+                "OpenAI model name lookups for LoRA adapters will fail.",
+                len(parsed),
+            )
+
        return adapters

    async def _ensure_engines_initialized(self):
@@ -242,16 +291,32 @@ class OpenAIvLLMEngine(vLLMEngine):
            lora_modules=self.lora_adapters,
        )
        await self.serving_models.init_static_loras()
-        
+
        # Get chat template from vLLM tokenizer if available
        chat_template = None
        if self.tokenizer and hasattr(self.tokenizer, 'tokenizer'):
            chat_template = self.tokenizer.tokenizer.chat_template
-        
+
+        self.openai_serving_render = OpenAIServingRender(
+            model_config=self.llm.model_config,
+            renderer=self.llm.renderer,
+            model_registry=self.serving_models.registry,
+            request_logger=None,
+            chat_template=chat_template,
+            chat_template_content_format="auto",
+            trust_request_chat_template=os.getenv('TRUST_REQUEST_CHAT_TEMPLATE', 'false').lower() == 'true',
+            enable_auto_tools=os.getenv('ENABLE_AUTO_TOOL_CHOICE', 'false').lower() == 'true',
+            exclude_tools_when_tool_choice_none=os.getenv('EXCLUDE_TOOLS_WHEN_TOOL_CHOICE_NONE', 'false').lower() == 'true',
+            tool_parser=os.getenv('TOOL_CALL_PARSER', "") or None,
+            reasoning_parser=os.getenv('REASONING_PARSER', "") or None,
+            log_error_stack=os.getenv('LOG_ERROR_STACK', 'false').lower() == 'true',
+        )
+
        self.chat_engine = OpenAIServingChat(
-            engine_client=self.llm, 
+            engine_client=self.llm,
            models=self.serving_models,
            response_role=self.response_role,
+            openai_serving_render=self.openai_serving_render,
            request_logger=None,
            chat_template=chat_template,
            chat_template_content_format="auto",
@@ -264,20 +329,53 @@ class OpenAIvLLMEngine(vLLMEngine):
            enable_prompt_tokens_details=os.getenv('ENABLE_PROMPT_TOKENS_DETAILS', 'false').lower() == 'true',
            enable_force_include_usage=os.getenv('ENABLE_FORCE_INCLUDE_USAGE', 'false').lower() == 'true',
            enable_log_outputs=os.getenv('ENABLE_LOG_OUTPUTS', 'false').lower() == 'true',
-            log_error_stack=os.getenv('LOG_ERROR_STACK', 'false').lower() == 'true',
        )
        self.completion_engine = OpenAIServingCompletion(
            engine_client=self.llm,
            models=self.serving_models,
+            openai_serving_render=self.openai_serving_render,
            request_logger=None,
            return_tokens_as_token_ids=os.getenv('RETURN_TOKENS_AS_TOKEN_IDS', 'false').lower() == 'true',
            enable_prompt_tokens_details=os.getenv('ENABLE_PROMPT_TOKENS_DETAILS', 'false').lower() == 'true',
            enable_force_include_usage=os.getenv('ENABLE_FORCE_INCLUDE_USAGE', 'false').lower() == 'true',
-            log_error_stack=os.getenv('LOG_ERROR_STACK', 'false').lower() == 'true',
+        )
+        self.responses_engine = OpenAIServingResponses(
+            engine_client=self.llm,
+            models=self.serving_models,
+            openai_serving_render=self.openai_serving_render,
+            request_logger=None,
+            chat_template=chat_template,
+            chat_template_content_format="auto",
+            return_tokens_as_token_ids=os.getenv('RETURN_TOKENS_AS_TOKEN_IDS', 'false').lower() == 'true',
+            reasoning_parser=os.getenv('REASONING_PARSER', "") or "",
+            enable_auto_tools=os.getenv('ENABLE_AUTO_TOOL_CHOICE', 'false').lower() == 'true',
+            tool_parser=os.getenv('TOOL_CALL_PARSER', "") or None,
+            tool_server=None,
+            enable_prompt_tokens_details=os.getenv('ENABLE_PROMPT_TOKENS_DETAILS', 'false').lower() == 'true',
+            enable_force_include_usage=os.getenv('ENABLE_FORCE_INCLUDE_USAGE', 'false').lower() == 'true',
+            enable_log_outputs=os.getenv('ENABLE_LOG_OUTPUTS', 'false').lower() == 'true',
+        )
+        self.messages_engine = AnthropicServingMessages(
+            engine_client=self.llm,
+            models=self.serving_models,
+            response_role=self.response_role,
+            openai_serving_render=self.openai_serving_render,
+            request_logger=None,
+            chat_template=chat_template,
+            chat_template_content_format="auto",
+            return_tokens_as_token_ids=os.getenv('RETURN_TOKENS_AS_TOKEN_IDS', 'false').lower() == 'true',
+            reasoning_parser=os.getenv('REASONING_PARSER', "") or "",
+            enable_auto_tools=os.getenv('ENABLE_AUTO_TOOL_CHOICE', 'false').lower() == 'true',
+            tool_parser=os.getenv('TOOL_CALL_PARSER', "") or None,
+            enable_prompt_tokens_details=os.getenv('ENABLE_PROMPT_TOKENS_DETAILS', 'false').lower() == 'true',
+            enable_force_include_usage=os.getenv('ENABLE_FORCE_INCLUDE_USAGE', 'false').lower() == 'true',
        )

-        if hasattr(self.chat_engine, 'warmup'):
-            await self.chat_engine.warmup()
+        warmup = getattr(self.chat_engine, 'warmup', None)
+        if callable(warmup):
+            result = warmup()
+            if inspect.isawaitable(result):
+                await result

    async def generate(self, openai_request: JobInput):
        # Ensure engines are ready (no-op if already initialized at startup)
@@ -288,6 +386,12 @@ class OpenAIvLLMEngine(vLLMEngine):
        elif openai_request.openai_route in ["/v1/chat/completions", "/v1/completions"]:
            async for response in self._handle_chat_or_completion_request(openai_request):
                yield response
+        elif openai_request.openai_route == "/v1/responses":
+            async for response in self._handle_responses_request(openai_request):
+                yield response
+        elif openai_request.openai_route == "/v1/messages":
+            async for response in self._handle_messages_request(openai_request):
+                yield response
        else:
            yield create_error_response("Invalid route").model_dump()
    
@@ -342,4 +446,127 @@ class OpenAIvLLMEngine(vLLMEngine):
                if self.raw_openai_output:
                    batch = "".join(batch)
                yield batch
-            
+
+    async def _handle_responses_request(self, openai_request: JobInput):
+        request_id = getattr(openai_request, "request_id", "unknown")
+
+        try:
+            request = ResponsesRequest(**openai_request.openai_input)
+        except Exception as e:
+            logging.error(
+                "Invalid ResponsesRequest JSON: %s",
+                e,
+                extra={"request_id": request_id}
+            )
+            yield create_error_response(
+                "Invalid request format",
+                err_type="BadRequestError"
+            ).model_dump()
+            return
+
+        dummy_request = DummyRequest()
+        try:
+            response = await self.responses_engine.create_responses(request, raw_request=dummy_request)
+        except Exception as e:
+            logging.error(
+                "Failed to create Responses: %s",
+                e,
+                extra={"request_id": request_id},
+                exc_info=True
+            )
+            yield create_error_response(
+                "Internal server error during response generation",
+                err_type="InternalServerError"
+            ).model_dump()
+            return
+
+        if isinstance(response, (ErrorResponse, ResponsesResponse)):
+            yield response.model_dump()
+            return
+
+        try:
+            async for event in response:
+                if not hasattr(event, "type"):
+                    continue
+                event_type = getattr(event, "type", "unknown")
+                yield f"event: {event_type}\ndata: {event.model_dump_json(indent=None)}\n\n"
+        except Exception as e:
+            logging.error(
+                "Error processing responses stream: %s",
+                e,
+                extra={"request_id": request_id},
+                exc_info=True
+            )
+            error_payload = create_error_response(
+                "Streaming response failed",
+                err_type="InternalServerError"
+            ).model_dump_json()
+            yield f"event: error\ndata: {error_payload}\n\n"
+
+    async def _handle_messages_request(self, openai_request: JobInput):
+        request_id = getattr(openai_request, "request_id", "unknown")
+
+        try:
+            request = AnthropicMessagesRequest(**openai_request.openai_input)
+        except Exception as e:
+            logging.error(
+                "Invalid AnthropicMessagesRequest: %s",
+                e,
+                extra={"request_id": request_id}
+            )
+            yield AnthropicErrorResponse(
+                error=AnthropicError(
+                    type="invalid_request_error",
+                    message="Invalid request format"
+                )
+            ).model_dump()
+            return
+
+        dummy_request = DummyRequest()
+
+        try:
+            response = await self.messages_engine.create_messages(request, raw_request=dummy_request)
+        except Exception as e:
+            logging.error(
+                "Failed to create messages: %s",
+                e,
+                extra={"request_id": request_id},
+                exc_info=True
+            )
+            yield AnthropicErrorResponse(
+                error=AnthropicError(
+                    type="internal_error",
+                    message="Failed to generate messages"
+                )
+            ).model_dump()
+            return
+
+        if isinstance(response, ErrorResponse):
+            error_type = getattr(response, "type", "internal_error")
+            error_message = getattr(response, "message", "Unknown error")
+            yield AnthropicErrorResponse(
+                error=AnthropicError(type=error_type, message=error_message)
+            ).model_dump()
+            return
+
+        if isinstance(response, AnthropicMessagesResponse):
+            yield response.model_dump(exclude_none=True)
+            return
+
+        try:
+            async for chunk in response:
+                yield chunk
+        except Exception as e:
+            logging.error(
+                "Error streaming messages: %s",
+                e,
+                extra={"request_id": request_id},
+                exc_info=True
+            )
+            error_payload = AnthropicErrorResponse(
+                error=AnthropicError(
+                    type="internal_error",
+                    message="Error while streaming messages"
+                )
+            ).model_dump_json()
+            yield f"event: error\ndata: {error_payload}\n\n"
@@ -1,106 +1,191 @@
+import ast
 import os
 import json
 import logging
+from typing import get_origin, get_args
 from torch.cuda import device_count
 from vllm import AsyncEngineArgs
 from vllm.model_executor.model_loader.tensorizer import TensorizerConfig
 from src.utils import convert_limit_mm_per_prompt

-RENAME_ARGS_MAP = {
+# Backward-compat: env var names users already know → engine arg name
+ENV_ALIASES = {
    "MODEL_NAME": "model",
    "MODEL_REVISION": "revision",
    "TOKENIZER_NAME": "tokenizer",
-    "MAX_CONTEXT_LEN_TO_CAPTURE": "max_seq_len_to_capture"
 }

+# Literal defaults from original worker (used when env/local do not set a value)
 DEFAULT_ARGS = {
-    "disable_log_stats": os.getenv('DISABLE_LOG_STATS', 'False').lower() == 'true',
-    # disable_log_requests is deprecated, use enable_log_requests instead
-    "enable_log_requests": os.getenv('ENABLE_LOG_REQUESTS', 'False').lower() == 'true',
-    "gpu_memory_utilization": float(os.getenv('GPU_MEMORY_UTILIZATION', 0.95)),
-    "pipeline_parallel_size": int(os.getenv('PIPELINE_PARALLEL_SIZE', 1)),
-    "tensor_parallel_size": int(os.getenv('TENSOR_PARALLEL_SIZE', 1)),
-    "served_model_name": os.getenv('SERVED_MODEL_NAME', None),
-    "tokenizer": os.getenv('TOKENIZER', None),
-    "skip_tokenizer_init": os.getenv('SKIP_TOKENIZER_INIT', 'False').lower() == 'true',
-    "tokenizer_mode": os.getenv('TOKENIZER_MODE', 'auto'),
-    "trust_remote_code": os.getenv('TRUST_REMOTE_CODE', 'False').lower() == 'true',
-    "download_dir": os.getenv('DOWNLOAD_DIR', None),
-    "load_format": os.getenv('LOAD_FORMAT', 'auto'),
-    "config_format": os.getenv('CONFIG_FORMAT', 'auto'),
-    "dtype": os.getenv('DTYPE', 'auto'),
-    "kv_cache_dtype": os.getenv('KV_CACHE_DTYPE', 'auto'),
-    "quantization_param_path": os.getenv('QUANTIZATION_PARAM_PATH', None),
-    "seed": int(os.getenv('SEED', 0)),
-    "max_model_len": int(os.getenv('MAX_MODEL_LEN', 0)) or None,
-    "worker_use_ray": os.getenv('WORKER_USE_RAY', 'False').lower() == 'true',
-    "distributed_executor_backend": os.getenv('DISTRIBUTED_EXECUTOR_BACKEND', None),
-    "max_parallel_loading_workers": int(os.getenv('MAX_PARALLEL_LOADING_WORKERS', 0)) or None,
-    "block_size": int(os.getenv('BLOCK_SIZE', 16)),
-    "enable_prefix_caching": os.getenv('ENABLE_PREFIX_CACHING', 'False').lower() == 'true',
-    "disable_sliding_window": os.getenv('DISABLE_SLIDING_WINDOW', 'False').lower() == 'true',
-    # attention_backend replaces deprecated VLLM_ATTENTION_BACKEND env var
-    "attention_backend": os.getenv('ATTENTION_BACKEND', None),
-    # Enabled by default for improved throughput. Set to False to disable if experiencing issues
-    "async_scheduling": None if os.getenv('ASYNC_SCHEDULING') is None else os.getenv('ASYNC_SCHEDULING', 'True').lower() == 'true',
-    # Controls how often to yield streaming results
-    "stream_interval": int(os.getenv('STREAM_INTERVAL', 1)),
-    "swap_space": int(os.getenv('SWAP_SPACE', 4)),  # GiB
-    "cpu_offload_gb": int(os.getenv('CPU_OFFLOAD_GB', 0)),  # GiB
-    # vLLM defaults None to 2048; keep 0 as None to let vLLM auto-calculate
-    "max_num_batched_tokens": int(os.getenv('MAX_NUM_BATCHED_TOKENS', 0)) or None,
-    "max_num_seqs": int(os.getenv('MAX_NUM_SEQS', 256)),
-    "max_logprobs": int(os.getenv('MAX_LOGPROBS', 20)),  # Default value for OpenAI Chat Completions API
-    "revision": os.getenv('REVISION', None),
-    "code_revision": os.getenv('CODE_REVISION', None),
-    "rope_scaling": os.getenv('ROPE_SCALING', None),
-    "rope_theta": float(os.getenv('ROPE_THETA', 0)) or None,
-    "tokenizer_revision": os.getenv('TOKENIZER_REVISION', None),
-    "quantization": os.getenv('QUANTIZATION', None),
-    "enforce_eager": os.getenv('ENFORCE_EAGER', 'False').lower() == 'true',
-    "max_context_len_to_capture": int(os.getenv('MAX_CONTEXT_LEN_TO_CAPTURE', 0)) or None,
-    "max_seq_len_to_capture": int(os.getenv('MAX_SEQ_LEN_TO_CAPTURE', 8192)),
-    "disable_custom_all_reduce": os.getenv('DISABLE_CUSTOM_ALL_REDUCE', 'False').lower() == 'true',
-    "tokenizer_pool_size": int(os.getenv('TOKENIZER_POOL_SIZE', 0)),
-    "tokenizer_pool_type": os.getenv('TOKENIZER_POOL_TYPE', 'ray'),
-    "tokenizer_pool_extra_config": os.getenv('TOKENIZER_POOL_EXTRA_CONFIG', None),
-    "enable_lora": os.getenv('ENABLE_LORA', 'False').lower() == 'true',
-    "max_loras": int(os.getenv('MAX_LORAS', 1)),
-    "max_lora_rank": int(os.getenv('MAX_LORA_RANK', 16)),
-    "enable_prompt_adapter": os.getenv('ENABLE_PROMPT_ADAPTER', 'False').lower() == 'true',
-    "max_prompt_adapters": int(os.getenv('MAX_PROMPT_ADAPTERS', 1)),
-    "max_prompt_adapter_token": int(os.getenv('MAX_PROMPT_ADAPTER_TOKEN', 0)),
-    "fully_sharded_loras": os.getenv('FULLY_SHARDED_LORAS', 'False').lower() == 'true',
-    "lora_extra_vocab_size": int(os.getenv('LORA_EXTRA_VOCAB_SIZE', 256)),
-    "long_lora_scaling_factors": tuple(map(float, os.getenv('LONG_LORA_SCALING_FACTORS', '').split(','))) if os.getenv('LONG_LORA_SCALING_FACTORS') else None,
-    "lora_dtype": os.getenv('LORA_DTYPE', 'auto'),
-    "max_cpu_loras": int(os.getenv('MAX_CPU_LORAS', 0)) or None,
-    "device": os.getenv('DEVICE', 'auto'),
-    "ray_workers_use_nsight": os.getenv('RAY_WORKERS_USE_NSIGHT', 'False').lower() == 'true',
-    "num_gpu_blocks_override": int(os.getenv('NUM_GPU_BLOCKS_OVERRIDE', 0)) or None,
-    "num_lookahead_slots": int(os.getenv('NUM_LOOKAHEAD_SLOTS', 0)),
-    "model_loader_extra_config": os.getenv('MODEL_LOADER_EXTRA_CONFIG', None),
-    "ignore_patterns": os.getenv('IGNORE_PATTERNS', None),
-    "preemption_mode": os.getenv('PREEMPTION_MODE', None),
-    "scheduler_delay_factor": float(os.getenv('SCHEDULER_DELAY_FACTOR', 0.0)),
-    "enable_chunked_prefill": os.getenv('ENABLE_CHUNKED_PREFILL', None),
-    "guided_decoding_backend": os.getenv('GUIDED_DECODING_BACKEND', 'outlines'),
-    "speculative_model": os.getenv('SPECULATIVE_MODEL', None),
-    "speculative_draft_tensor_parallel_size": int(os.getenv('SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE', 0)) or None,
-    "enable_expert_parallel": bool(os.getenv('ENABLE_EXPERT_PARALLEL', 'False').lower() == 'true'),
-    "num_speculative_tokens": int(os.getenv('NUM_SPECULATIVE_TOKENS', 0)) or None,
-    "speculative_max_model_len": int(os.getenv('SPECULATIVE_MAX_MODEL_LEN', 0)) or None,
-    "speculative_disable_by_batch_size": int(os.getenv('SPECULATIVE_DISABLE_BY_BATCH_SIZE', 0)) or None,
-    "ngram_prompt_lookup_max": int(os.getenv('NGRAM_PROMPT_LOOKUP_MAX', 0)) or None,
-    "ngram_prompt_lookup_min": int(os.getenv('NGRAM_PROMPT_LOOKUP_MIN', 0)) or None,
-    "spec_decoding_acceptance_method": os.getenv('SPEC_DECODING_ACCEPTANCE_METHOD', 'rejection_sampler'),
-    "typical_acceptance_sampler_posterior_threshold": float(os.getenv('TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_THRESHOLD', 0)) or None,
-    "typical_acceptance_sampler_posterior_alpha": float(os.getenv('TYPICAL_ACCEPTANCE_SAMPLER_POSTERIOR_ALPHA', 0)) or None,
-    "qlora_adapter_name_or_path": os.getenv('QLORA_ADAPTER_NAME_OR_PATH', None),
-    "disable_logprobs_during_spec_decoding": os.getenv('DISABLE_LOGPROBS_DURING_SPEC_DECODING', None),
-    "otlp_traces_endpoint": os.getenv('OTLP_TRACES_ENDPOINT', None),
+    "disable_log_stats": False,
+    "enable_log_requests": False,
+    "gpu_memory_utilization": 0.95,
+    "pipeline_parallel_size": 1,
+    "tensor_parallel_size": 1,
+    "skip_tokenizer_init": False,
+    "tokenizer_mode": "auto",
+    "trust_remote_code": False,
+    "load_format": "auto",
+    "dtype": "auto",
+    "kv_cache_dtype": "auto",
+    "seed": 0,
+    "worker_use_ray": False,
+    "block_size": 16,
+    "enable_prefix_caching": False,
+    "disable_sliding_window": False,
+    "swap_space": 4,
+    "cpu_offload_gb": 0,
+    "max_num_seqs": 256,
+    "max_logprobs": 20,
+    "enforce_eager": False,
+    "max_seq_len_to_capture": 8192,
+    "disable_custom_all_reduce": False,
+    "tokenizer_pool_size": 0,
+    "tokenizer_pool_type": "ray",
+    "enable_lora": False,
+    "max_loras": 1,
+    "max_lora_rank": 16,
+    "enable_prompt_adapter": False,
+    "max_prompt_adapters": 1,
+    "max_prompt_adapter_token": 0,
+    "fully_sharded_loras": False,
+    "lora_extra_vocab_size": 256,
+    "lora_dtype": "auto",
+    "device": "auto",
+    "ray_workers_use_nsight": False,
+    "num_lookahead_slots": 0,
+    "scheduler_delay_factor": 0.0,
+    "guided_decoding_backend": "outlines",
+    "spec_decoding_acceptance_method": "rejection_sampler",
+    "stream_interval": 1,
+
 }

+
+def _resolve_field_type(field_type: type) -> type:
+    """Resolve Optional/Union to the concrete type for conversion."""
+    origin = get_origin(field_type)
+    args = get_args(field_type) if hasattr(field_type, "__args__") else ()
+    if origin is not None:
+        # Optional[X] is Union[X, None]; X | None is UnionType
+        non_none = [a for a in args if a is not type(None)]
+        if non_none:
+            return non_none[0]
+    return field_type
+
+
+def _convert_env_value_to_field_type(value: str, field_name: str, field_type: type):
+    """Convert env var string to the type expected by AsyncEngineArgs for this field."""
+    val = value.strip() if isinstance(value, str) else value
+    if val in ("", "None", "none"):
+        args = get_args(field_type) if hasattr(field_type, "__args__") else ()
+        if type(None) in (args or ()):
+            return None
+        raise ValueError("empty value not allowed for non-optional field")
+
+    # Union[bool, str, ...]: only coerce to bool for unambiguous literals;
+    # otherwise preserve the string (e.g. hf_token="hf_abc..." must stay a str).
+    if get_origin(field_type) is not None:
+        union_types = [a for a in (get_args(field_type) or ()) if a is not type(None)]
+        if bool in union_types and str in union_types:
+            if str(val).lower() in ("true", "false", "1", "0", "yes", "no", "on", "off"):
+                return str(val).lower() in ("true", "1", "yes", "on")
+            return str(val)
+
+    effective_type = _resolve_field_type(field_type)
+    # bool
+    if effective_type is bool:
+        return str(val).lower() in ("true", "1", "yes", "on")
+    # int
+    if effective_type is int:
+        return int(val)
+    # float
+    if effective_type is float:
+        return float(val)
+    # str
+    if effective_type is str:
+        return str(val)
+    # dict, list, or complex (try JSON)
+    origin = get_origin(effective_type)
+    if effective_type in (dict, list) or origin in (dict, list):
+        try:
+            return json.loads(val)
+        except json.JSONDecodeError:
+            return val
+    # tuple (e.g. long_lora_scaling_factors) — comma-separated or JSON array
+    if effective_type is tuple or origin is tuple:
+        args = get_args(field_type) if hasattr(field_type, "__args__") else ()
+        elem_types = [a for a in args if a is not Ellipsis]
+        elem_type = elem_types[0] if elem_types else str
+        try:
+            parsed = json.loads(val)
+            if isinstance(parsed, list):
+                return tuple(elem_type(x) for x in parsed)
+        except (json.JSONDecodeError, TypeError):
+            pass
+        return tuple(elem_type(x.strip()) for x in str(val).split(",") if x.strip())
+    # For dataclass/complex types, try JSON then Python literal parsing to dict
+    try:
+        return json.loads(val)
+    except (json.JSONDecodeError, TypeError):
+        pass
+    try:
+        parsed = ast.literal_eval(val)
+        if isinstance(parsed, (dict, list)):
+            return parsed
+    except (ValueError, SyntaxError):
+        pass
+    # Fallback: try int, float, then str
+    try:
+        return int(val)
+    except ValueError:
+        pass
+    try:
+        return float(val)
+    except ValueError:
+        pass
+    return str(val)
+
+
+def _get_args_from_env_auto_discover() -> dict:
+    """Auto-discover engine args from env vars using UPPERCASED field names.
+
+    For every field in AsyncEngineArgs, check os.getenv(FIELD_NAME).
+    E.g. MAX_MODEL_LEN=4096 -> max_model_len=4096.
+    Uses same type conversion as before; supports all vLLM engine args without manual listing.
+    """
+    args = {}
+    valid_fields = AsyncEngineArgs.__dataclass_fields__
+    for field_name, field in valid_fields.items():
+        env_key = field_name.upper()
+        value = os.environ.get(env_key)
+        if value is None:
+            continue
+        try:
+            args[field_name] = _convert_env_value_to_field_type(
+                value, field_name, field.type
+            )
+        except (ValueError, TypeError, json.JSONDecodeError) as e:
+            logging.warning(
+                "Skip env %s=%r: %s", env_key, value, e
+            )
+    return args
+
+
+def _apply_env_aliases(args: dict) -> None:
+    """Apply ENV_ALIASES: if MODEL_NAME etc. are set, set the target engine arg."""
+    valid_fields = AsyncEngineArgs.__dataclass_fields__
+    for alias, target in ENV_ALIASES.items():
+        value = os.environ.get(alias)
+        if value is None or target not in valid_fields:
+            continue
+        try:
+            args[target] = _convert_env_value_to_field_type(
+                value, target, valid_fields[target].type
+            )
+        except (ValueError, TypeError, json.JSONDecodeError) as e:
+            logging.warning("Skip env alias %s=%r: %s", alias, value, e)
+
 def get_speculative_config():
    """Build speculative decoding configuration from environment variables.

@@ -122,9 +207,14 @@ def get_speculative_config():
    # Option 2: Build config from individual environment variables
    spec_method = os.getenv('SPECULATIVE_METHOD')
    spec_model = os.getenv('SPECULATIVE_MODEL')
-    num_spec_tokens = os.getenv('NUM_SPECULATIVE_TOKENS')
-    ngram_max = os.getenv('NGRAM_PROMPT_LOOKUP_MAX')
-    ngram_min = os.getenv('NGRAM_PROMPT_LOOKUP_MIN')
+    _num_spec_tokens = os.getenv('NUM_SPECULATIVE_TOKENS')
+    _ngram_max = os.getenv('NGRAM_PROMPT_LOOKUP_MAX')
+    _ngram_min = os.getenv('NGRAM_PROMPT_LOOKUP_MIN')
+
+    # Convert numeric vars to int so '0' (hub.json default) is treated as unset
+    num_spec_tokens = (int(_num_spec_tokens) or None) if _num_spec_tokens else None
+    ngram_max = (int(_ngram_max) or None) if _ngram_max else None
+    ngram_min = (int(_ngram_min) or None) if _ngram_min else None

    if not any([spec_method, spec_model, ngram_max]):
        return None
@@ -150,11 +240,11 @@ def get_speculative_config():
    if spec_model:
        config['model'] = spec_model
    if num_spec_tokens:
-        config['num_speculative_tokens'] = int(num_spec_tokens)
+        config['num_speculative_tokens'] = num_spec_tokens
    if ngram_max:
-        config['prompt_lookup_max'] = int(ngram_max)
+        config['prompt_lookup_max'] = ngram_max
    if ngram_min:
-        config['prompt_lookup_min'] = int(ngram_min)
+        config['prompt_lookup_min'] = ngram_min

    draft_tp = os.getenv('SPECULATIVE_DRAFT_TENSOR_PARALLEL_SIZE')
    if draft_tp:
@@ -186,6 +276,7 @@ def get_speculative_config():

    return None

+
 def _resolve_max_model_len(model, trust_remote_code=False, revision=None):
    """Resolve max_model_len from the model's HuggingFace config."""
    try:
@@ -204,25 +295,132 @@ def _resolve_max_model_len(model, trust_remote_code=False, revision=None):
        logging.warning(f"Could not resolve max_model_len from model config: {e}")
    return None

-limit_mm_env = os.getenv('LIMIT_MM_PER_PROMPT')
-if limit_mm_env is not None:
-    DEFAULT_ARGS["limit_mm_per_prompt"] = convert_limit_mm_per_prompt(limit_mm_env)

-def match_vllm_args(args):
-    """Rename args to match vllm by:
-    1. Renaming keys to lower case
-    2. Renaming keys to match vllm
-    3. Filtering args to match vllm's AsyncEngineArgs
+def _local_args_to_engine_args(local: dict) -> dict:
+    """Map local args (e.g. from /local_model_args.json) to engine arg names and filter."""
+    valid = AsyncEngineArgs.__dataclass_fields__
+    out = {}
+    for k, v in local.items():
+        target = ENV_ALIASES.get(k, k.lower().replace("-", "_"))
+        if target not in valid or v in (None, "", "None"):
+            continue
+        out[target] = v
+    return out

-    Args:
-        args (dict): Dictionary of args

-    Returns:
-        dict: Dictionary of args with renamed keys
+def _sanitize_hf_overrides(hf_overrides: dict) -> dict | None:
+    """Strip rope_scaling from hf_overrides sub-configs if vLLM rejects them.
+
+    Older vLLM (<0.7) required explicit mrope rope_scaling in hf_overrides for
+    models like Qwen2-VL. Newer vLLM auto-detects mrope and raises a ValueError
+    in patch_rope_scaling_dict when it finds conflicting rope_type values. Strip
+    the offending rope_scaling so the model loads with its native config.
    """
-    renamed_args = {RENAME_ARGS_MAP.get(k, k): v for k, v in args.items()}
-    matched_args = {k: v for k, v in renamed_args.items() if k in AsyncEngineArgs.__dataclass_fields__}
-    return {k: v for k, v in matched_args.items() if v not in [None, "", "None"]}
+    if not isinstance(hf_overrides, dict):
+        return hf_overrides
+
+    try:
+        from vllm.transformers_utils.config import patch_rope_scaling_dict
+    except ImportError:
+        return hf_overrides
+
+    import copy
+    cleaned = {}
+    changed = False
+    for key, value in hf_overrides.items():
+        if isinstance(value, dict) and "rope_scaling" in value:
+            rope_scaling = value.get("rope_scaling")
+            if isinstance(rope_scaling, dict):
+                try:
+                    patch_rope_scaling_dict(copy.deepcopy(rope_scaling))
+                except (ValueError, Exception) as e:
+                    logging.warning(
+                        "Stripping hf_overrides['%s']['rope_scaling'] because vLLM "
+                        "rejected it (%s). Newer vLLM auto-detects rope scaling from "
+                        "the model config.", key, e
+                    )
+                    stripped = {k: v for k, v in value.items() if k != "rope_scaling"}
+                    cleaned[key] = stripped if stripped else None
+                    changed = True
+                    continue
+        cleaned[key] = value
+
+    if not changed:
+        return hf_overrides
+
+    result = {k: v for k, v in cleaned.items() if v is not None}
+    return result or None
+
+
+def _resolve_cached_model_path(model_name: str) -> str:
+    """Return a local snapshot path when the HF cache was stored with lowercase names.
+
+    Some model stores (e.g. RunPod pre-cached volumes) normalize repo IDs to
+    lowercase. HuggingFace Hub stores caches as
+    ``models--{org}--{model}/snapshots/{hash}/`` preserving the original casing,
+    so MODEL_NAME=Qwen/Qwen2.5-Coder-32B-Instruct-AWQ will miss a cache stored
+    as ``models--qwen--qwen2.5-coder-32b-instruct-awq/``.
+
+    If the exact-case cache directory is absent but a lowercase variant exists,
+    the latest snapshot path is returned so vLLM loads from disk rather than
+    attempting a redundant download.
+    """
+    if os.path.isabs(model_name):
+        return model_name
+
+    cache_dir = (
+        os.getenv("HUGGINGFACE_HUB_CACHE")
+        or os.getenv("HF_HOME")
+        or os.path.expanduser("~/.cache/huggingface/hub")
+    )
+
+    folder_name = f"models--{model_name.replace('/', '--')}"
+
+    if os.path.isdir(os.path.join(cache_dir, folder_name)):
+        return model_name
+
+    lower_dir = os.path.join(cache_dir, folder_name.lower())
+    if not os.path.isdir(lower_dir):
+        return model_name
+
+    snapshots_dir = os.path.join(lower_dir, "snapshots")
+    if not os.path.isdir(snapshots_dir):
+        return model_name
+
+    try:
+        snapshots = sorted(os.listdir(snapshots_dir))
+    except OSError:
+        return model_name
+
+    if not snapshots:
+        return model_name
+
+    resolved = os.path.join(snapshots_dir, snapshots[-1])
+    logging.info(
+        "MODEL_NAME %r not found at original casing in HF cache; "
+        "resolved to lowercase cached snapshot at %r",
+        model_name, resolved,
+    )
+    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.
@@ -245,23 +443,46 @@ def get_local_args():

    return local_args
 def get_engine_args():
-    # Start with default args
-    args = DEFAULT_ARGS
-    
-    # Get env args that match keys in AsyncEngineArgs
-    args.update(os.environ)
-    
-    # Get local args if model is baked in and overwrite env args
-    args.update(get_local_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())
+
+    # Backward-compat aliases (MODEL_NAME → model, etc.)
+    _apply_env_aliases(args)
+
+    # Local baked-in model overrides
+    local = get_local_args()
+    if local:
+        args.update(_local_args_to_engine_args(local))
+
+    # Filter to valid engine args and drop sentinel empty values
+    valid_fields = AsyncEngineArgs.__dataclass_fields__
+    args = {
+        k: v for k, v in args.items()
+        if k in valid_fields and v not in (None, "", "None")
+    }
+
+    # Special conversion for limit_mm_per_prompt (e.g. "image=1,video=0")
+    limit_mm_env = os.getenv("LIMIT_MM_PER_PROMPT")
+    if limit_mm_env is not None:
+        args["limit_mm_per_prompt"] = convert_limit_mm_per_prompt(limit_mm_env)
+
    # if args.get("TENSORIZER_URI"): TODO: add back once tensorizer is ready
    #     args["load_format"] = "tensorizer"
    #     args["model_loader_extra_config"] = TensorizerConfig(tensorizer_uri=args["TENSORIZER_URI"], num_readers=None)
    #     logging.info(f"Using tensorized model from {args['TENSORIZER_URI']}")
-    
-    
-    # Rename and match to vllm args
-    args = match_vllm_args(args)
+
+    if "hf_overrides" in args:
+        sanitized = _sanitize_hf_overrides(args["hf_overrides"])
+        if sanitized:
+            args["hf_overrides"] = sanitized
+        else:
+            del args["hf_overrides"]

    if args.get("load_format") == "bitsandbytes":
        args["quantization"] = args["load_format"]
@@ -274,6 +495,53 @@ def get_engine_args():
        if os.getenv("MAX_PARALLEL_LOADING_WORKERS"):
            logging.warning("Overriding MAX_PARALLEL_LOADING_WORKERS with None because more than 1 GPU is available.")
    
+    # LMCache requires HMA to be disabled
+    try:
+        _kv_transfer = args.get("kv_transfer_config")
+        if isinstance(_kv_transfer, str):
+            parsed = None
+            try:
+                parsed = json.loads(_kv_transfer)
+            except (json.JSONDecodeError, TypeError):
+                pass
+            if parsed is None:
+                try:
+                    result = ast.literal_eval(_kv_transfer)
+                    if isinstance(result, dict):
+                        parsed = result
+                except (ValueError, SyntaxError):
+                    pass
+            if parsed is not None:
+                _kv_transfer = parsed
+                args["kv_transfer_config"] = _kv_transfer
+        _kv_offload = args.get("kv_offloading_backend")
+
+        lmcache_via_offload = _kv_offload == "lmcache"
+        lmcache_via_transfer = (
+            isinstance(_kv_transfer, dict)
+            and isinstance(_kv_transfer.get("kv_connector"), str)
+            and "lmcache" in _kv_transfer.get("kv_connector", "").lower()
+        )
+        lmcache_detected = lmcache_via_offload or lmcache_via_transfer
+
+        if lmcache_detected:
+            current = args.get("disable_hybrid_kv_cache_manager")
+            if current is False:
+                logging.warning(
+                    "disable_hybrid_kv_cache_manager=False conflicts with LMCache; "
+                    "overriding to True (HMA must be disabled when using LMCache)"
+                )
+                args["disable_hybrid_kv_cache_manager"] = True
+            elif current is None:
+                args["disable_hybrid_kv_cache_manager"] = True
+                logging.info("LMCache detected: automatically setting disable_hybrid_kv_cache_manager=True")
+    except Exception as e:
+        logging.error(
+            "Failed to check LMCache configuration: %s",
+            e,
+            exc_info=True
+        )
+
    # Deprecated env args backwards compatibility
    if args.get("kv_cache_dtype") == "fp8_e5m2":
        args["kv_cache_dtype"] = "fp8"
@@ -331,4 +599,8 @@ def get_engine_args():
    if speculative_config:
        args["speculative_config"] = speculative_config

+    # Resolve lowercase HF cache paths (FDE-174)
+    if args.get("model"):
+        args["model"] = _resolve_cached_model_path(args["model"])
+
    return AsyncEngineArgs(**args)
@@ -0,0 +1,9 @@
+#!/bin/bash
+set -e
+
+if [ -n "${TRANSFORMERS_VERSION}" ]; then
+    echo "Installing transformers==${TRANSFORMERS_VERSION}"
+    uv pip install --system "transformers==${TRANSFORMERS_VERSION}"
+fi
+
+exec python3 /src/handler.py
@@ -1,10 +1,12 @@
-from transformers import AutoTokenizer
+import logging
 import os
 from typing import Union

+from transformers import AutoTokenizer
+
 class TokenizerWrapper:
    def __init__(self, tokenizer_name_or_path, tokenizer_revision, trust_remote_code):
-        print(f"tokenizer_name_or_path: {tokenizer_name_or_path}, tokenizer_revision: {tokenizer_revision}, trust_remote_code: {trust_remote_code}")
+        logging.debug("tokenizer_name_or_path: %s, tokenizer_revision: %s, trust_remote_code: %s", tokenizer_name_or_path, tokenizer_revision, trust_remote_code)
        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path, revision=tokenizer_revision or "main", trust_remote_code=trust_remote_code)
        self.custom_chat_template = os.getenv("CUSTOM_CHAT_TEMPLATE")
        self.has_chat_template = bool(self.tokenizer.chat_template) or bool(self.custom_chat_template)
Author	SHA1	Message	Date
chrisvelaandGitHub	d3a962c33b	Merge pull request #301 from adithyaJRunpod/feature/tuned-configs Release / release (push) Waiting to run Add tuned configs, CON-239	2026-06-11 14:26:00 -05:00
chrisvelaandGitHub	105c125698	Merge pull request #300 from runpod-workers/feat/0.21.0 feat: upgrade vllm to 0.21.0	2026-06-11 12:47:08 -05:00
AdithyaJob	0a0ccfcb60	Add tuned configs for Llama 3.1 8B and Qwen3 8B	2026-06-10 21:07:41 -07:00
velaraptor-runpod	c8ce53c72c	fix: add kenels, and fix for cuda	2026-06-10 16:01:53 -05:00
velaraptor-runpod	9618e799ba	chore: fix cuda libraries	2026-06-04 15:35:58 -05:00
velaraptor-runpod	cb3f077dba	feat: upgrade vllm to 0.21.0	2026-06-03 14:43:56 -05:00
chrisvelaandGitHub	69646b9e99	Merge pull request #294 from runpod-workers/feat/allow-config Release / release (push) Waiting to run feat: allow config.yaml like vllm serve	2026-06-02 20:28:02 -05:00
Jacob CiparandGitHub	8b991a7ad7	Merge pull request #297 from runpod-workers/jhcipar/bump-runpod-python-version feat: bump runpod-python version	2026-06-02 11:17:48 -04:00
Tim PietruskyandGitHub	dac05b62b3	fix: make .runpod/tests.json hub tests pass on CUDA 13.0 (#299 ) Release / release (push) Waiting to run Three coupled fixes verified end-to-end on a private fork (TimPietruskyRunPod/worker-vllm v0.1.3 → both hub tests passing): 1. tests.json allowedCudaVersions: 12.x → 13.0 The Dockerfile and hub.json moved to CUDA 13.0 in v2.20.0 (#288, #289), but tests.json was still pinned to 12.5–12.9, so the test pod was scheduled on a GPU with driver < 13.0 and container init failed at the nvidia-container-cli hook with "unsatisfied condition: cuda>=13.0". 2. requirements.txt kernels<0.15 huggingface/kernels v0.15.1 tightened LayerRepository to require a revision or version argument (https://github.com/huggingface/kernels/pull/544). transformers >=5 still constructs LayerRepository(repo_id=..., layer_name=...) without either, so worker import raised ValueError during `from transformers import ...`. 0.14.1 is the last safe release. 3. tests.json timeout 30000 → 300000 vLLM cold start (torch.compile + FlashInfer warmup) on RTX 4090 for SmolLM2-135M takes ~60–70s before the first request can be served. The previous 30s per-test timeout fired before the worker came up, producing "context cancelled or timed out: context deadline exceeded" for every test even when the worker was healthy. 300s gives enough headroom for cold start + the actual inference call. Refs: DR-1161	2026-06-02 17:08:54 +02:00
jhcipar	d356c31675	feat: bump runpod-python version	2026-06-01 20:28:40 -04:00
Tim PietruskyandGitHub	14b74a4989	chore: re-enable .runpod/tests.json hub tests (#295 ) Release / release (push) Waiting to run Rename tests_json back to tests.json to re-enable the automated hub tests that were temporarily disabled in #253. Refs: DR-1161	2026-06-01 17:52:23 +02:00
velaraptor-runpod	80072047ab	feat: allow config.yaml like vllm serve	2026-05-29 15:50:22 -05:00
chrisvelaandGitHub	50aba8fb57	Merge pull request #293 from runpod-workers/fix/update-deep-gemm-hub-value Release / release (push) Waiting to run fix: update VLLM_USE_DEEP_GEMM hub to default to 0	2026-05-27 11:51:48 -05:00
velaraptor-runpod	9edc5715ce	fix: update configuration.md	2026-05-27 11:33:37 -05:00
velaraptor-runpod	4c91f2c5b5	fix: update VLLM_USE_DEEP_GEMM hub to default to 0	2026-05-27 11:26:52 -05:00
chrisvelaandGitHub	6265b99348	Merge pull request #288 from runpod-workers/feat/0.20.0 feat: update to 0.20.2	2026-05-26 17:17:45 -05:00
velaraptor-runpod	026f8d700b	fix: specify deepgemm commit version	2026-05-21 18:46:54 -05:00
chrisvelaandGitHub	146bdb0252	Merge branch 'main' into feat/0.20.0	2026-05-21 16:04:39 -05:00
velaraptor-runpod	da01193a3d	chore: fix readme	2026-05-21 15:57:40 -05:00
velaraptor-runpod	c2e6cc9f61	chore: update readme with correct vllm version	2026-05-21 15:39:19 -05:00
velaraptor-runpod	69968a6b39	chore: fix logging, warning for text prompt	2026-05-21 15:34:09 -05:00
velaraptor-runpod	32b29d4c6c	fix: add deepgemm, update base image and hub for cuda 13.0	2026-05-20 17:43:42 -05:00
velaraptor-runpod	dcea4fc4f9	fix dockerfile	2026-05-15 12:08:31 -04:00
velaraptor-runpod	9c139e8ceb	update: update to 0.20.1, update dockerfile to cuda 13	2026-05-15 11:35:53 -04:00
velaraptor-runpod	678bb4be8f	feat: update to 0.20.1 for patch fixes	2026-05-07 11:58:22 -05:00
chrisvelaandGitHub	87d7365126	Merge pull request #292 from runpod-workers/fix/open-ai Release / release (push) Waiting to run fix: fix warmup	2026-05-01 18:06:11 -05:00
velaraptor-runpod	0e83616f93	fix: fix warmup	2026-05-01 17:39:51 -05:00
chrisvelaandGitHub	ab6d39dcf8	Merge pull request #291 from runpod-workers/bug/281-hf-token Release / release (push) Waiting to run bug: fix hf-token being passed in engineargs	2026-05-01 15:32:04 -05:00
velaraptor-runpod	ed315a175e	merge main	2026-05-01 15:28:05 -05:00
velaraptor-runpod	73f030ae5e	Merge branch 'main' into bug/281-hf-token	2026-05-01 14:33:51 -05:00
chrisvelaandGitHub	8a099c1723	Merge pull request #287 from runpod-workers/feat/0.19.1 feat: update to 0.19.1	2026-05-01 14:29:39 -05:00
velaraptor-runpod	ff87840a58	fix: add enforce_eager as true, add pytorch_alloc_conf to expandle_segments to True for OOM, for hub defaults	2026-05-01 14:13:03 -05:00
velaraptor-runpod	7dc853b1fe	chore: remove release to trigger on release publish, just use tags. duplicate	2026-05-01 10:13:32 -05:00
velaraptor-runpod	a8b754b92a	merge main	2026-05-01 10:12:43 -05:00
chrisvelaandGitHub	6357aeda51	Merge pull request #290 from runpod-workers/fix/fix-old-actions Release / release (push) Waiting to run fix: fix old github actions, trigger release on publish release	2026-05-01 10:01:22 -05:00
velaraptor-runpod	0140b29c44	chore: add release notes to slack notification	2026-05-01 09:48:43 -05:00
velaraptor-runpod	0cb8aeae77	chore: add specific runpod version	2026-05-01 09:47:23 -05:00
chrisvelaandGitHub	cd8f9e9560	Merge branch 'main' into fix/fix-old-actions	2026-04-30 21:57:45 -05:00
chrisvelaandGitHub	895fd25fac	Merge pull request #286 from runpod-workers/feat/0.18.1 feat: update vllm to 0.18.1	2026-04-30 21:56:07 -05:00
velaraptor-runpod	7bb8df73af	bug: fix hf-token being passed in engineargs	2026-04-30 20:37:30 -05:00
velaraptor-runpod	747cdf5891	chore: update readme vllm version	2026-04-30 20:26:23 -05:00
velaraptor-runpod	3d4af5df9b	chore: update readme vllm version	2026-04-30 20:25:35 -05:00
velaraptor-runpod	72547aa3bb	chore: update readme vllm version	2026-04-30 20:25:02 -05:00
velaraptor-runpod	577fd8c3c3	fix: fix old github actions, trigger release on publish release	2026-04-30 20:22:24 -05:00
chrisvelaandGitHub	f49f35456e	Merge pull request #285 from runpod-workers/feat/0.17.1 feat: update vllm to 0.17.1	2026-04-30 20:05:45 -05:00
chrisvelaandGitHub	cff7b09ef2	Merge pull request #289 from runpod-workers/feat/add-notifications feat: add notifications for new prs, issues, and new releases of vllm	2026-04-30 20:03:48 -05:00
velaraptor-runpod	04b342c675	fix permissions	2026-04-30 20:00:48 -05:00
velaraptor-runpod	5b29643799	feat: add notifications for new prs, issues, and new releases of vllm	2026-04-30 19:55:29 -05:00
velaraptor-runpod	4f8a16df5d	fix: update transformers to >=5	2026-04-30 19:41:09 -05:00
velaraptor-runpodandClaude Sonnet 4.6	22356ee2b3	feat: upgrade vLLM to 0.20.0 - Bump vllm[flashinfer] to 0.20.0 in Dockerfile - Remove io_processor param from OpenAIServingRender (dropped in 0.20.0) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-30 18:39:07 -05:00
velaraptor-runpodandClaude Sonnet 4.6	fa42ecd79a	fix: resolve lowercase HF cache paths when MODEL_NAME uses original casing Fixes FDE-174. Some model stores (e.g. RunPod pre-cached network volumes) normalize repo IDs to lowercase. HuggingFace Hub caches using the original casing, so MODEL_NAME=Qwen/Qwen2.5-Coder-32B-Instruct-AWQ would miss a cache stored as models--qwen--qwen2.5-coder-32b-instruct-awq/ and attempt a redundant download that fails on limited container storage. If the exact-case HF cache directory is absent but a lowercase variant exists, the latest snapshot path is returned directly so vLLM loads from disk. Absolute paths and models with no lowercase cache are unchanged. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-30 17:54:14 -05:00
velaraptor-runpodandClaude Sonnet 4.6	178c72238e	fix: surface LORA_MODULES parse failures instead of silently loading zero adapters Fixes FDE-194. Previously a malformed LORA_MODULES value was swallowed at info level and the engine would start with no LoRA adapters, causing 500s on any request using an adapter model name (e.g. npc-sim-*). Changes: - Log at error level when LORA_MODULES cannot be parsed as JSON - Log at error level when individual adapter dicts fail LoRAModulePath validation - Log a final error when all adapters fail to load so the cause is obvious - Accept a single adapter dict (not just an array) for convenience - Return early when LORA_MODULES is unset to skip unnecessary parsing Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-30 17:49:09 -05:00
velaraptor-runpodandClaude Sonnet 4.6	e6950bdebd	feat: upgrade vLLM to 0.19.1 - Bump vllm[flashinfer] to 0.19.1 in Dockerfile - Add OpenAIServingRender (new required dependency in 0.19.x serving layer) - Pass openai_serving_render to all four serving class constructors - Remove log_error_stack param (removed upstream in 0.19.x) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>	2026-04-30 17:41:11 -05:00
velaraptor-runpod	a774cefe85	fix: clean workspace file	2026-04-30 17:29:00 -05:00
velaraptor-runpod	9de17d49b7	feat: update vllm to 0.18.1	2026-04-30 17:25:23 -05:00
velaraptor-runpod	296556a6f7	feat: update vllm to 0.17.1	2026-04-30 17:21:03 -05:00
chrisvelaandGitHub	a1544ea70d	Merge pull request #277 from runpod-workers/feat/lmcache feat: uv installer, LMCache support, add /v1/responses and /v1/messages endpoints	2026-04-28 19:44:00 -03:00
Tim Pietrusky	3403889528	fix: address review comments on responses/messages handlers and lmcache guard - engine.py: drop UnboundLocalError-prone isinstance(response, ...) checks in except blocks of _handle_responses_request and _handle_messages_request; emit SSE-shaped error frames mid-stream instead of raw dicts; add missing blank line between handlers. - engine_args.py: restructure LMCache HMA guard so the warning branch is actually reachable when user explicitly sets disable_hybrid_kv_cache_manager=False, and correct the inverted message (HMA must be disabled = True). - requirements.txt: drop stray whitespace in transformers version specifier.	2026-04-23 11:28:50 +02:00
Tim Pietrusky	f299204770	docs: fix anthropic messages path and missing comma in routes list	2026-04-23 10:54:24 +02:00
velaraptor-runpod	1ed25eea20	update readme with responses and messages routes	2026-04-16 15:30:30 -05:00
velaraptor-runpod	dc4ad7ddeb	fix: kv_transfer_config is dataclass not json, fix for env var	2026-04-09 14:33:11 -05:00
velaraptor-runpod	9035b0e07f	fix: lmcache version	2026-04-09 14:32:19 -05:00
velaraptor-runpod	c979f0020f	update readmes on TRANSFORMERS_VERSION	2026-04-06 16:54:08 -05:00
velaraptor-runpod	6fbd480a26	fix: allow for transformers_version	2026-04-06 16:49:32 -05:00
velaraptor-runpod	3ef1fb8e7b	requested changes	2026-04-03 15:53:40 -05:00
velaraptor-runpod	30e8514d63	Runpod not RunPod	2026-03-17 20:59:52 -05:00
velaraptor-runpod	d9808815ee	feat: update requirements.txt	2026-03-17 19:56:39 -05:00
velaraptor-runpod	d8ed3b5353	feat: update 0.16.0, add lmcache	2026-03-17 19:54:11 -05:00
velaraptor-runpod	4c4e039565	feat: add messages route for anthropic/claude	2026-03-17 19:51:50 -05:00
chrisvelaandGitHub	9d1686960d	Merge pull request #273 from runpod-workers/bug/hf-overides-rope-scaling bug: fix rope scaling to be forward compatible from hf_overrides	2026-03-10 11:21:44 -05:00
velaraptor-runpod	45d1eeee47	bug: fix rope scaling to be forward compatible from hf_overrides	2026-03-06 15:34:11 -06:00
chrisvelaandGitHub	17efb0e7d0	Merge pull request #272 from runpod-workers/feat/vllm-0.16.0 Release / release (push) Waiting to run feat: Update to 0.16.0	2026-03-05 13:06:45 -06:00
velaraptor-runpod	2b5f07df63	feat: Update to 0.16.0, remove NUM_GPU_BLOCKS_OVERRIDE in hub default since 0 will break	2026-03-04 16:38:40 -06:00
chrisvelaandGitHub	13fa71878e	Merge pull request #269 from runpod-workers/feat/allow-engine-args-env Release / release (push) Waiting to run feat: allow all AsyncEngineArgs as env vars	2026-02-27 15:34:23 -06:00
velaraptor-runpod	8a9365bed4	remove DEFAULT_ARGS that are none, fix MAX_CONTEXT_LEN_TO_CAPTURE	2026-02-27 14:04:15 -06:00
velaraptor-runpod	cd485a1af1	update readme	2026-02-25 22:57:17 -06:00
velaraptor-runpod	b9043639e9	requested changes/refactor	2026-02-25 16:07:38 -06:00
chrisvelaandGitHub	407dbd7773	Merge pull request #270 from runpod-workers/feat/update-vllm-v0.15.1 feat: update vllm to 0.15.1	2026-02-25 15:35:58 -06:00
velaraptor-runpod	f103c142c1	feat: update vllm to 0.15.1	2026-02-24 17:44:49 -06:00
velaraptor-runpod	efb093e198	add as VLLM_RUNPOD prefix and update readme	2026-02-24 17:37:57 -06:00
velaraptor-runpod	42443f735e	feat: allow engine args through VLLM_ and checks the engine args	2026-02-24 16:05:18 -06:00
chrisvelaandGitHub	b7c6d4f9a2	feat: update dockerfile to 12.9.1 (#267 ) Release / release (push) Waiting to run * feat: update dockerfile to 12.9.1 * update readme on VLLM_NIGHTLY build arg	2026-02-19 10:13:14 +01:00
chrisvelaandGitHub	d69cc021e8	Merge pull request #268 from runpod-workers/fix/spec-config-0-to-none Release / release (push) Waiting to run fix: spec config env vars should be none if zero	2026-02-18 15:51:51 -06:00
velaraptor-runpod	61faa8f137	fix: spec config env vars should be none if zero	2026-02-18 15:41:19 -06:00