Unsloth chat template for qwen3.5

Fix shell quoting for exec command arguments

Dockerfile

@@ -113,7 +113,9 @@ RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     cd flashinfer-cubin && uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v && \
     # flashinfer-jit-cache
     cd ../flashinfer-jit-cache && \
-    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v
+    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v && \
+    # dump git ref in the wheels dir
+    cd .. && git rev-parse HEAD > /workspace/wheels/.flashinfer-commit
 
 # =========================================================
 # STAGE 3: FlashInfer Wheel Export
@@ -196,7 +198,9 @@ RUN curl -L https://patch-diff.githubusercontent.com/raw/vllm-project/vllm/pull/
 # Final Compilation
 RUN --mount=type=cache,id=ccache,target=/root/.ccache \
     --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v
+    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v && \
+    # dump git ref in the wheels dir
+    git rev-parse HEAD > /workspace/wheels/.vllm-commit
 
 # =========================================================
 # STAGE 5: vLLM Wheel Export

README.md

@@ -26,7 +26,10 @@ While it was primarily developed to support multi-node inference, it works just
 
 This repository is not affiliated with NVIDIA or their subsidiaries. This is a community effort aimed to help DGX Spark users to set up and run the most recent versions of vLLM on Spark cluster or single nodes. 
 
-The Dockerfile builds from the main branch of VLLM, so depending on when you run the build process, it may not be in fully functioning state. You can target a specific vLLM release by setting `--vllm-ref` parameter.
+Unless `--rebuild-vllm` or `--vllm-ref` or `--apply-vllm-pr` is specified, the builder will fetch the latest precompiled vLLM wheels from the repository. They are built nightly and tested on multiple models in both cluster and solo configuration before publishing.
+We will expand the selection of models we test in the pipeline, but since vLLM is a rapidly developing platform, some things may break.
+
+If you want to build the latest from main branch, you can specify `--rebuild-vllm` flag. Or you can target a specific vLLM release by setting `--vllm-ref` parameter.
 
 ## QUICK START
 
@@ -58,7 +61,7 @@ Then run the following command that will build and distribute image across the c
 ./build-and-copy.sh -c
 ```
 
-An initial build will take around 20-30 minutes, but subsequent builds will be faster. Precompiled vLLM wheels for DGX Spark will also be available soon.
+An initial build speed depends on your Internet connection speed and whether the base image is already present on your machine. After base image pull, the build should take only 2-3 minutes. If `--rebuild-vllm` and/or `--rebuild-flashinfer` is used to trigger a build from the sourcew, it will take between 20-40 minutes, but subsequent builds will be faster. Prebuilt FlashInfer and vLLM wheels are downloaded automatically from GitHub releases, so compilation from source is usually not required.
 
 ### Run
 
@@ -120,7 +123,7 @@ To launch the model:
 
 This will run the model on all available cluster nodes.
 
-**NOTE:** do not use `--load-format fastsafetensors` if you are loading models that would take >0.8 of available RAM (without KV cache) as it may result in out of memory situation.
+**NOTE:** do not use `--load-format fastsafetensors` if you are loading models that would take >0.85 of available RAM (without KV cache) as it may result in out of memory situation.
 
 **Also:** You can use any vLLM container that has "bash" as its default entrypoint with the launch script. It was tested with NGC vLLM, but can work with others too. To use such container in the cluster, you need to specify `--apply-mod use-ngc-vllm` argument to `./launch-cluster.sh`. However, it's recommended to build the container using this repository for best compatibility and most up-to-date features. 
 
@@ -146,6 +149,21 @@ Don't do it every time you rebuild, because it will slow down compilation times.
 
 For periodic maintenance, I recommend using a filter: `docker builder prune --filter until=72h`
 
+### 2026-03-04
+
+#### Prebuilt vLLM Wheels via GitHub Releases
+
+`build-and-copy.sh` now automatically downloads prebuilt vLLM wheels from the [GitHub releases](https://github.com/eugr/spark-vllm-docker/releases/tag/prebuilt-vllm-current) before falling back to a local build — identical to the existing FlashInfer download mechanism. This eliminates the need to compile vLLM from source on first use.
+
+The download logic mirrors the FlashInfer behaviour:
+- If prebuilt wheels are available and newer than any locally cached version, they are downloaded automatically.
+- If the download fails (e.g. no network, release not found, GPU arch not supported), the script falls back to building locally, or reuses existing local wheels if present.
+- `--rebuild-vllm`, `--vllm-ref`, or `--apply-vllm-pr` skip the download entirely and force a local build.
+
+No new flags are required — the download happens transparently.
+
+All prebuilt wheels are now tested with multiple models in both solo and cluster configuration as a part of automated deployment pipeline which will now run nightly. The wheels are released only if they pass all the tests and no significant performance regressions are detected. 
+
 ### 2026-03-02
 
 #### Qwen3.5-122B-INT4-Autoround Support
@@ -178,7 +196,6 @@ Added a new mod for Intel/Qwen3-Coder-Next-INT4-Autoround model support: `mods/f
 
 Changed reasoning parser in Minimax for better compatibility with modern clients (like coding tools).
 
-
 ### 2026-02-18
 
 #### Completely Redesigned Build Process

build-and-copy.sh

@@ -23,6 +23,7 @@ BUILD_JOBS="16"
 GPU_ARCH_LIST="12.1a"
 WHEELS_REPO="eugr/spark-vllm-docker"
 FLASHINFER_RELEASE_TAG="prebuilt-flashinfer-current"
+VLLM_RELEASE_TAG="prebuilt-vllm-current"
 # Space-separated list of GPU architectures for which prebuilt wheels are available
 PREBUILT_WHEELS_SUPPORTED_ARCHS="12.1a"
 
@@ -65,7 +66,12 @@ copy_to_host() {
 
 # try_download_wheels TAG PREFIX
 # Downloads wheels matching PREFIX*.whl from a GitHub release.
-# Skips files that are already present and up to date (by remote updated_at vs local mtime).
+# Skip conditions (either is sufficient):
+#   1. Commit hash in release name matches .wheels/.{PREFIX}_commit (primary check).
+#   2. All local wheels are newer than the latest GitHub asset (freshly built).
+# Only downloads a file when the remote asset is newer than the local copy AND
+# the above skip conditions are not met.
+# On success, persists the release commit hash to .wheels/.{PREFIX}_commit.
 # Returns 0 if all matching wheels are now available, 1 on any error.
 try_download_wheels() {
     local TAG="$1"
@@ -91,7 +97,7 @@ try_download_wheels() {
 
     local DOWNLOAD_LIST
     DOWNLOAD_LIST=$(echo "$RELEASE_JSON" | python3 -c '
-import json, sys, os
+import json, sys, os, re
 from datetime import datetime, timezone
 
 wheels_dir, prefix = sys.argv[1], sys.argv[2]
@@ -103,6 +109,31 @@ if not assets:
     print("No assets found matching prefix: " + prefix, file=sys.stderr)
     sys.exit(1)
 
+# Extract commit hash from the release name:
+#   FlashInfer: "Prebuilt FlashInfer Wheels (0.6.5-124a2d32-d20260305) - DGX Spark Only"
+#   vLLM:       "Prebuilt vLLM Wheels (0.16.1rc1.dev296+ga73af584f.d20260305.cu131) - DGX Spark only"
+release_name = data.get("name", "")
+commit_hash = None
+if prefix.startswith("flashinfer"):
+    m = re.search(r"\([\d.]+\w*-([0-9a-f]{6,})-d\d{8}\)", release_name, re.IGNORECASE)
+    if m:
+        commit_hash = m.group(1)
+else:
+    m = re.search(r"\+g([0-9a-f]{6,})\.", release_name, re.IGNORECASE)
+    if m:
+        commit_hash = m.group(1)
+
+# Compare against the locally stored commit hash
+commit_file = os.path.join(wheels_dir, "." + prefix + "-commit")
+local_commit = None
+if os.path.exists(commit_file):
+    with open(commit_file) as f:
+        local_commit = f.read().strip()
+
+if commit_hash and local_commit and local_commit[:len(commit_hash)] == commit_hash:
+    print("Commit hash matches (" + commit_hash + ") — wheels are up to date.", file=sys.stderr)
+    sys.exit(0)
+
 newest_remote_ts = max(
     datetime.strptime(a["updated_at"], "%Y-%m-%dT%H:%M:%SZ")
             .replace(tzinfo=timezone.utc).timestamp()
@@ -118,12 +149,19 @@ local_wheels = [
 if local_wheels and all(os.path.getmtime(p) >= newest_remote_ts for p in local_wheels):
     sys.exit(0)
 
+downloads = []
 for a in assets:
     local_path = os.path.join(wheels_dir, a["name"])
     remote_ts = datetime.strptime(a["updated_at"], "%Y-%m-%dT%H:%M:%SZ") \
                     .replace(tzinfo=timezone.utc).timestamp()
     if not os.path.exists(local_path) or remote_ts > os.path.getmtime(local_path):
-        print(a["browser_download_url"] + " " + a["name"])
+        downloads.append(a["browser_download_url"] + " " + a["name"])
+
+if downloads:
+    if commit_hash:
+        print("#commit:" + commit_hash)
+    for d in downloads:
+        print(d)
 ' "$WHEELS_DIR" "$PREFIX") || return 1
 
     if [ -z "$DOWNLOAD_LIST" ]; then
@@ -131,12 +169,31 @@ for a in assets:
         return 0
     fi
 
+    # Parse the optional '#commit:HASH' sentinel emitted by the Python script
+    local REMOTE_COMMIT=""
+    local DOWNLOAD_ENTRIES=""
+    while IFS= read -r LINE; do
+        if [[ "$LINE" == "#commit:"* ]]; then
+            REMOTE_COMMIT="${LINE#"#commit:"}"
+        elif [[ -n "$LINE" ]]; then
+            DOWNLOAD_ENTRIES+="$LINE"$'\n'
+        fi
+    done <<< "$DOWNLOAD_LIST"
+
+    if [ -z "$DOWNLOAD_ENTRIES" ]; then
+        echo "All $PREFIX wheels are up to date — skipping download."
+        return 0
+    fi
+
     # Back up existing wheels so we never leave a mix of old and new on failure
     local DL_BACKUP="$WHEELS_DIR/.backup-download-${PREFIX}"
     rm -rf "$DL_BACKUP" && mkdir -p "$DL_BACKUP"
     for f in "$WHEELS_DIR/${PREFIX}"*.whl; do
         [ -f "$f" ] && mv "$f" "$DL_BACKUP/"
     done
+    for f in "$WHEELS_DIR/.${PREFIX}"*; do
+        [ -f "$f" ] && mv "$f" "$DL_BACKUP/"
+    done
 
     local URL NAME TMP_WHL
     local DOWNLOADED=()
@@ -153,13 +210,18 @@ for a in assets:
             if compgen -G "$DL_BACKUP/${PREFIX}*.whl" > /dev/null 2>&1; then
                 echo "Restoring previous $PREFIX wheels..."
                 mv "$DL_BACKUP/${PREFIX}"*.whl "$WHEELS_DIR/"
+                mv "$DL_BACKUP/.${PREFIX}"* "$WHEELS_DIR/"
             fi
             rm -rf "$DL_BACKUP"
             return 1
         fi
-    done <<< "$DOWNLOAD_LIST"
+    done <<< "$DOWNLOAD_ENTRIES"
 
     rm -rf "$DL_BACKUP"
+    if [ -n "$REMOTE_COMMIT" ]; then
+        echo "$REMOTE_COMMIT" > "$WHEELS_DIR/.${PREFIX}-commit"
+        echo "Recorded $PREFIX commit hash: $REMOTE_COMMIT"
+    fi
     return 0
 }
 
@@ -347,30 +409,32 @@ if [ "$NO_BUILD" = false ]; then
         # ----------------------------------------------------------
         # Phase 2: vLLM wheels
         # ----------------------------------------------------------
-        VLLM_WHEELS_EXIST=false
-        if compgen -G "./wheels/vllm*.whl" > /dev/null 2>&1; then
-            VLLM_WHEELS_EXIST=true
-        fi
-
         if [ "$VLLM_REF_SET" = true ] || [ -n "$VLLM_PRS" ]; then
             REBUILD_VLLM=true
         fi
 
-        if [ "$REBUILD_VLLM" = true ] || [ "$VLLM_WHEELS_EXIST" = false ]; then
-            if [ "$REBUILD_VLLM" = true ]; then
-                if [ "$VLLM_REF_SET" = true ] && [ -n "$VLLM_PRS" ]; then
-                    echo "Rebuilding vLLM wheels (--vllm-ref and --apply-vllm-pr specified)..."
-                elif [ "$VLLM_REF_SET" = true ]; then
-                    echo "Rebuilding vLLM wheels (--vllm-ref specified)..."
-                elif [ -n "$VLLM_PRS" ]; then
-                    echo "Rebuilding vLLM wheels (--apply-vllm-pr specified)..."
-                else
-                    echo "Rebuilding vLLM wheels (--rebuild-vllm specified)..."
-                fi
+        BUILD_VLLM=false
+        if [ "$REBUILD_VLLM" = true ]; then
+            if [ "$VLLM_REF_SET" = true ] && [ -n "$VLLM_PRS" ]; then
+                echo "Rebuilding vLLM wheels (--vllm-ref and --apply-vllm-pr specified)..."
+            elif [ "$VLLM_REF_SET" = true ]; then
+                echo "Rebuilding vLLM wheels (--vllm-ref specified)..."
+            elif [ -n "$VLLM_PRS" ]; then
+                echo "Rebuilding vLLM wheels (--apply-vllm-pr specified)..."
             else
-                echo "No vLLM wheels found in ./wheels/ — building..."
+                echo "Rebuilding vLLM wheels (--rebuild-vllm specified)..."
             fi
+            BUILD_VLLM=true
+        elif try_download_wheels "$VLLM_RELEASE_TAG" "vllm"; then
+            echo "vLLM wheels ready."
+        elif compgen -G "./wheels/vllm*.whl" > /dev/null 2>&1; then
+            echo "Download failed — using existing local vLLM wheels."
+        else
+            echo "No vLLM wheels available (download failed) — building..."
+            BUILD_VLLM=true
+        fi
 
+        if [ "$BUILD_VLLM" = true ]; then
             # Back up existing vllm wheels; restore them if the build fails
             VLLM_BACKUP="./wheels/.backup-vllm"
             rm -rf "$VLLM_BACKUP" && mkdir -p "$VLLM_BACKUP"
@@ -393,7 +457,6 @@ if [ "$NO_BUILD" = false ]; then
                 VLLM_CMD+=("--build-arg" "VLLM_PRS=$VLLM_PRS")
             fi
 
-
             VLLM_CMD+=(".")
 
             echo "vLLM build command: ${VLLM_CMD[*]}"
@@ -408,8 +471,6 @@ if [ "$NO_BUILD" = false ]; then
                 rm -rf "$VLLM_BACKUP"
                 exit 1
             fi
-        else
-            echo "vLLM wheels already present in ./wheels/ — skipping build."
         fi
 
         # ----------------------------------------------------------

docs/NETWORKING.md

@@ -58,8 +58,8 @@ network:
   ethernets:
     enp1s0f1np1:
       dhcp4: no
-      dhcp6: no              # Explicitly disable DHCPv6
-      link-local: [ ipv4 ]   # Restrict link-local addresses to IPv4 only
+      dhcp6: no        # Explicitly disable DHCPv6
+      link-local: []   # Restrict link-local addresses to static IPv4 only
       mtu: 9000
       addresses: [192.168.177.11/24]
     enP2p1s0f1np1:
@@ -76,8 +76,8 @@ network:
   ethernets:
     enp1s0f1np1:
       dhcp4: no
-      dhcp6: no              # Explicitly disable DHCPv6
-      link-local: [ ipv4 ]   # Restrict link-local addresses to IPv4 only
+      dhcp6: no        # Explicitly disable DHCPv6
+      link-local: []   # Restrict link-local addresses to static IPv4 only
       mtu: 9000
       addresses: [192.168.177.12/24]
     enP2p1s0f1np1:
@@ -239,4 +239,4 @@ mpirun -np 2 -H 192.168.177.11:1,192.168.177.12:1 \
   -x LD_LIBRARY_PATH=$LD_LIBRARY_PATH \
   $HOME/nccl-tests/build/all_gather_perf -b 16G -e 16G -f 2
 
-```
+```

launch-cluster.sh

@@ -115,7 +115,7 @@ while [[ "$#" -gt 0 ]]; do
             fi
             ACTION="exec"
             shift
-            COMMAND_TO_RUN="$@"
+            COMMAND_TO_RUN=$(printf "%q " "$@")
             break
             ;;
         *) 

mods/fix-qwen3.5-chat-template/chat_template.jinja

@@ -0,0 +1,155 @@
+{%- set image_count = namespace(value=0) %}
+{%- set video_count = namespace(value=0) %}
+{%- macro render_content(content, do_vision_count, is_system_content=false) %}
+    {%- if content is string %}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping %}
+        {%- for item in content %}
+            {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain images.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set image_count.value = image_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Picture ' ~ image_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
+            {%- elif 'video' in item or item.type == 'video' %}
+                {%- if is_system_content %}
+                    {{- raise_exception('System message cannot contain videos.') }}
+                {%- endif %}
+                {%- if do_vision_count %}
+                    {%- set video_count.value = video_count.value + 1 %}
+                {%- endif %}
+                {%- if add_vision_id %}
+                    {{- 'Video ' ~ video_count.value ~ ': ' }}
+                {%- endif %}
+                {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
+            {%- elif 'text' in item %}
+                {{- item.text }}
+            {%- else %}
+                {{- raise_exception('Unexpected item type in content.') }}
+            {%- endif %}
+        {%- endfor %}
+    {%- elif content is none or content is undefined %}
+        {{- '' }}
+    {%- else %}
+        {{- raise_exception('Unexpected content type.') }}
+    {%- endif %}
+{%- endmacro %}
+{%- if not messages %}
+    {{- raise_exception('No messages provided.') }}
+{%- endif %}
+{%- if tools and tools is iterable and tools is not mapping %}
+    {{- '<|im_start|>system\n' }}
+    {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>" }}
+    {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {%- if content %}
+            {{- '\n\n' + content }}
+        {%- endif %}
+    {%- endif %}
+    {{- '<|im_end|>\n' }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {%- set content = render_content(messages[0].content, false, true)|trim %}
+        {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" %}
+        {%- set content = render_content(message.content, false)|trim %}
+        {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
+            {%- set ns.multi_step_tool = false %}
+            {%- set ns.last_query_index = index %}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if ns.multi_step_tool %}
+    {{- raise_exception('No user query found in messages.') }}
+{%- endif %}
+{%- for message in messages %}
+    {%- set content = render_content(message.content, true)|trim %}
+    {%- if message.role == "system" %}
+        {%- if not loop.first %}
+            {{- raise_exception('System message must be at the beginning.') }}
+        {%- endif %}
+    {%- elif message.role == "user" %}
+        {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is string %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in content %}
+                {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+                {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- set reasoning_content = reasoning_content|trim %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if tool_call.function is defined %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {%- if loop.first %}
+                    {%- if content|trim %}
+                        {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- else %}
+                        {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                    {%- endif %}
+                {%- else %}
+                    {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
+                {%- endif %}
+                {%- if tool_call.arguments is mapping %}
+                    {%- for args_name in tool_call.arguments %}
+                        {%- set args_value = tool_call.arguments[args_name] %}
+                        {{- '<parameter=' + args_name + '>\n' }}
+                        {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
+                        {{- args_value }}
+                        {{- '\n</parameter>\n' }}
+                    {%- endfor %}
+                {%- endif %}
+                {{- '</function>\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.previtem and loop.previtem.role != "tool" %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- content }}
+        {{- '\n</tool_response>' }}
+        {%- if not loop.last and loop.nextitem.role != "tool" %}
+            {{- '<|im_end|>\n' }}
+        {%- elif loop.last %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- else %}
+        {{- raise_exception('Unexpected message role.') }}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- else %}
+        {{- '<think>\n' }}
+    {%- endif %}
+{%- endif %}

mods/fix-qwen3.5-chat-template/run.sh

@@ -0,0 +1,4 @@
+#!/bin/bash
+set -e
+cp chat_template.jinja $WORKSPACE_DIR/unsloth.jinja
+echo "=======> to apply chat template, use --chat-template unsloth.jinja"

recipes/4x-spark-cluster/minimax-m2.5.yaml

@@ -0,0 +1,45 @@
+# Recipe: MiniMax-M2.5
+# MiniMaxAI/MiniMax-M2.5
+
+recipe_version: "1"
+name: MiniMax-M2.5
+description: vLLM serving MiniMax-M2.5 with Ray distributed backend
+
+# HuggingFace model to download (optional, for --download-model)
+model: MiniMaxAI/MiniMax-M2.5
+
+# Container image to use
+container: vllm-node
+
+# Can only be run in a cluster
+cluster_only: true
+
+# No mods required
+mods: []
+
+# Default settings (can be overridden via CLI)
+defaults:
+  port: 8000
+  host: 0.0.0.0
+  tensor_parallel: 4
+  gpu_memory_utilization: 0.90
+  max_model_len: 128000
+
+# Environment variables
+env:
+  VLLM_DISTRIBUTED_EXECUTOR_CONFIG: '{"placement_group_options":{"strategy":"SPREAD"}}'
+
+# The vLLM serve command template
+command: |
+  vllm serve MiniMaxAI/MiniMax-M2.5 \
+      --trust-remote-code \
+      --port {port} \
+      --host {host} \
+      --gpu-memory-utilization {gpu_memory_utilization} \
+      -tp {tensor_parallel} \
+      --distributed-executor-backend ray \
+      --max-model-len {max_model_len} \
+      --load-format fastsafetensors \
+      --enable-auto-tool-choice \
+      --tool-call-parser minimax_m2 \
+      --reasoning-parser minimax_m2_append_think

recipes/4x-spark-cluster/qwen3.5-397b-a17B-fp8.yaml

@@ -0,0 +1,63 @@
+# Recipe: Qwen3.5-397B-A17B-FP8
+# Qwen3.5-397B-A17B model in FP8 precision
+# Multi-modal input
+
+recipe_version: "1"
+name: Qwen3.5-397B-A17B-FP8
+description: vLLM serving Qwen3.5-397B-A17B-FP8
+
+# HuggingFace model to download (optional, for --download-model)
+model: Qwen/Qwen3.5-397B-A17B-FP8
+
+#solo_only: true
+
+# Container image to use
+container: vllm-node-tf5
+
+build_args:
+  - --tf5
+  - --rebuild-flashinfer
+  - --rebuild-vllm
+
+# Mod required to fix ROPE syntax error
+mods:
+  - mods/fix-qwen3.5-autoround
+
+# Default settings (can be overridden via CLI)
+defaults:
+  port: 8000
+  host: 0.0.0.0
+  tensor_parallel: 4
+  gpu_memory_utilization: 0.85
+  max_model_len: 262144
+  max_num_batched_tokens: 8192
+
+# Environment variables
+env: 
+  VLLM_USE_DEEP_GEMM: 0
+  VLLM_USE_FLASHINFER_MOE_FP16: 1
+  VLLM_USE_FLASHINFER_SAMPLER: 0
+  OMP_NUM_THREADS: 4
+
+# The vLLM serve command template
+command: |
+  vllm serve Qwen/Qwen3.5-397B-A17B-FP8 \
+    --max-model-len {max_model_len} \
+    --gpu-memory-utilization {gpu_memory_utilization} \
+    --port {port} \
+    --host {host} \
+    --load-format fastsafetensors \
+    --enable-prefix-caching \
+    --enable-auto-tool-choice \
+    --tool-call-parser qwen3_coder \
+    --reasoning-parser qwen3 \
+    --max-num-batched-tokens {max_num_batched_tokens} \
+    --trust-remote-code \
+    -tp {tensor_parallel} \
+    --distributed-executor-backend ray \
+    --mm-encoder-tp-mode data \
+    --kv-cache-dtype fp8 \
+    --compilation-config.cudagraph_mode none \
+    --max-num-seqs 32 \
+    --attention-backend flashinfer
+

recipes/qwen3.5-122b-fp8.yaml

@@ -0,0 +1,47 @@
+# Recipe: Qwen3.5-122B-A10B-FP8
+# Qwen3.5-122B model in native FP8 quantization
+
+recipe_version: "1"
+name: Qwen3.5-122B-FP8
+description: vLLM serving Qwen3.5-122B-FP8
+
+# HuggingFace model to download (optional, for --download-model)
+model: Qwen/Qwen3.5-122B-A10B-FP8
+
+# Only cluster is supported
+cluster_only: true
+
+# Container image to use
+container: vllm-node
+
+# No mods required
+mods:
+  - mods/fix-qwen3.5-chat-template
+
+# Default settings (can be overridden via CLI)
+defaults:
+  port: 8000
+  host: 0.0.0.0
+  tensor_parallel: 2
+  gpu_memory_utilization: 0.7
+  max_model_len: 262144
+  max_num_batched_tokens: 8192
+
+# Environment variables
+env: {}
+
+# The vLLM serve command template
+command: |
+  vllm serve Qwen/Qwen3.5-122B-A10B-FP8 \
+    --max-model-len {max_model_len} \
+    --gpu-memory-utilization {gpu_memory_utilization} \
+    --port {port} \
+    --host {host} \
+    --load-format fastsafetensors \
+    --enable-prefix-caching \
+    --enable-auto-tool-choice \
+    --tool-call-parser qwen3_coder \
+    --reasoning-parser qwen3 \
+    --chat-template unsloth.jinja \
+    -tp {tensor_parallel} --distributed-executor-backend ray \
+    --max-num-batched-tokens {max_num_batched_tokens}

recipes/qwen3.5-122b-int4-autoround.yaml

@@ -19,6 +19,7 @@ build_args:
 # Mod required to fix ROPE syntax error
 mods:
   - mods/fix-qwen3.5-autoround
+  - mods/fix-qwen3.5-chat-template
 
 # Default settings (can be overridden via CLI)
 defaults:
@@ -47,6 +48,7 @@ command: |
     --reasoning-parser qwen3 \
     --max-num-batched-tokens {max_num_batched_tokens} \
     --trust-remote-code \
+    --chat-template unsloth.jinja \
     -tp {tensor_parallel} \
     --distributed-executor-backend ray
 

recipes/qwen3.5-35b-a3b-fp8.yaml

@@ -0,0 +1,51 @@
+# Recipe: Qwen/Qwen3.5-35B-A3B-FP8
+# Qwen/Qwen3.5-35B-A3B model in native FP8 format
+
+
+recipe_version: "1"
+name: Qwen35-35B-A3B
+description: vLLM serving Qwen3.5-35B-A3B-FP8
+
+# HuggingFace model to download (optional, for --download-model)
+model: Qwen/Qwen3.5-35B-A3B-FP8
+
+#solo_only: true
+
+# Container image to use
+container: vllm-node
+
+# Mod required to fix slowness and crash in the cluster (tracking https://github.com/vllm-project/vllm/issues/33857)
+mods:
+  - mods/fix-qwen3-coder-next
+  - mods/fix-qwen3.5-chat-template
+
+# Default settings (can be overridden via CLI)
+defaults:
+  port: 8000
+  host: 0.0.0.0
+  tensor_parallel: 2
+  gpu_memory_utilization: 0.7
+  max_model_len: 262144
+  max_num_batched_tokens: 16384
+
+# Environment variables
+env: 
+  VLLM_MARLIN_USE_ATOMIC_ADD: 1
+
+# The vLLM serve command template
+command: |
+  vllm serve Qwen/Qwen3.5-35B-A3B-FP8 \
+    --host {host} \
+    --port {port} \
+    --max-model-len {max_model_len} \
+    --max-num-batched-tokens {max_num_batched_tokens} \
+    --gpu-memory-utilization {gpu_memory_utilization} \
+    --enable-auto-tool-choice \
+    --tool-call-parser qwen3_coder \
+    --kv-cache-dtype fp8 \
+    --load-format fastsafetensors \
+    --attention-backend flashinfer \
+    --enable-prefix-caching \
+    --chat-template unsloth.jinja \
+    -tp {tensor_parallel} \
+    --distributed-executor-backend ray

run-recipe.py

@@ -86,6 +86,7 @@ RELATED FILES:
 import argparse
 import os
 import subprocess
+import shlex
 import sys
 import tempfile
 from pathlib import Path
@@ -473,7 +474,7 @@ def generate_launch_script(recipe: dict[str, Any], overrides: dict[str, Any], is
     # Append extra args if provided (after --)
     if extra_args:
         # Join extra args and append to command
-        extra_args_str = ' '.join(extra_args)
+        extra_args_str = ' '.join(shlex.quote(a) for a in extra_args)
         command = command.rstrip()
         # Handle multi-line commands with backslash continuations
         if command.endswith('\\'):