Merge branch 'mxfp4'

Dockerfile.mxfp4

@@ -99,7 +99,7 @@ ARG FLASHINFER_SHA=f349e52496a72a00d8c4ac02c7a1e38523ff7194
 ARG CUTLASS_SHA=11af7f02ab52c9130e422eeb4b44042fbd60c083
 
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-     uv pip install nvidia-nvshmem-cu13
+     uv pip install nvidia-nvshmem-cu13 "apache-tvm-ffi<0.2"
 
 # Clone FlashInfer (cached for faster rebuilds)
 RUN --mount=type=cache,id=git-flashinfer,target=/git-cache/flashinfer \
@@ -144,18 +144,18 @@ WORKDIR /workspace/flashinfer
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     --mount=type=cache,id=ccache,target=/root/.ccache \
     sed -i -e 's/license = "Apache-2.0"/license = { text = "Apache-2.0" }/' -e '/license-files/d' pyproject.toml && \
-    uv pip install --no-build-isolation . -v
+    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v
 
 # flashinfer-cubin
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     --mount=type=cache,id=ccache,target=/root/.ccache \
-    cd flashinfer-cubin && uv pip install --no-build-isolation . -v
+    cd flashinfer-cubin && uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v
 
 # flashinfer-jit-cache
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
     --mount=type=cache,id=ccache,target=/root/.ccache \
     cd flashinfer-jit-cache && \
-    uv pip install --no-build-isolation . -v
+    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v
 
 # --- VLLM SOURCE CACHE BUSTER ---
 # Change THIS argument to force a fresh git clone and rebuild of vLLM
@@ -218,7 +218,7 @@ RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
 # across totally separate `docker build` invocations.
 RUN --mount=type=cache,id=ccache,target=/root/.ccache \
     --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv pip install --no-build-isolation . -v
+    uv build --no-build-isolation --wheel . --out-dir=/workspace/wheels -v
 
 # =========================================================
 # STAGE 4: Runner (Transfers only necessary artifacts)
@@ -256,8 +256,9 @@ RUN mkdir -p tiktoken_encodings && \
 # Copy artifacts from Builder Stage
 # We copy the python packages and executables
 # No need to copy source code, as it's already in the site-packages
-COPY --from=builder /usr/local/lib/python3.12/dist-packages /usr/local/lib/python3.12/dist-packages
-COPY --from=builder /usr/local/bin /usr/local/bin
+COPY --from=builder /workspace/wheels /workspace/wheels
+RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
+    uv pip install /workspace/wheels/*.whl
 
 # Setup Env for Runtime
 ENV TORCH_CUDA_ARCH_LIST="12.0;12.1"
@@ -272,4 +273,4 @@ RUN chmod +x $VLLM_BASE_DIR/run-cluster-node.sh
 
 # Final extra deps
 RUN --mount=type=cache,id=uv-cache,target=/root/.cache/uv \
-    uv pip install ray[default]
+    uv pip install ray[default] fastsafetensors

README.md

@@ -146,9 +146,11 @@ For periodic maintenance, I recommend using a filter: `docker builder prune --fi
 
 ### 2026-01-29
 
+Added `-e` / `--env` parameter to `launch-cluster.sh` to pass environment variables to the container.
+
 Added an experimental build option, optimized for DGX Spark and gpt-oss models by [Christopher Owen](https://github.com/christopherowen/spark-vllm-mxfp4-docker/blob/main/Dockerfile).
 
-It is currently the fastest way to run GPT-OSS on DGX Spark, achieving 60 t/s on a single Spark and 75 t/s on dual Sparks. 
+It is currently the fastest way to run GPT-OSS on DGX Spark, achieving 60 t/s on a single Spark.
 
 To use this build, first build the container with `--exp-mxfp4` flag. I recommend using a separate label as it is currently not recommended to use this build for models other than gpt-oss:
 
@@ -521,7 +523,7 @@ The script attempts to automatically detect:
 You can override the auto-detected values if needed:
 
 ```bash
-./launch-cluster.sh --nodes "10.0.0.1,10.0.0.2" --eth-if enp1s0f1np1 --ib-if rocep1s0f1
+./launch-cluster.sh --nodes "10.0.0.1,10.0.0.2" --eth-if enp1s0f1np1 --ib-if rocep1s0f1 -e MY_ENV=123
 ```
 
 | Flag | Description |
@@ -531,6 +533,7 @@ You can override the auto-detected values if needed:
 | `--name` | Container name (default: `vllm_node`). |
 | `--eth-if` | Ethernet interface name. |
 | `--ib-if` | InfiniBand interface name. |
+| `-e, --env` | Environment variable to pass to container (e.g. `-e VAR=val`). Can be used multiple times. |
 | `--apply-mod` | Apply mods/patches from specified directory. Can be used multiple times to apply multiple mods. |
 | `--nccl-debug` | NCCL debug level (e.g., INFO, WARN). Defaults to INFO if flag is present but value is omitted. |
 | `--check-config` | Check configuration and auto-detection without launching. |

launch-cluster.sh

@@ -35,6 +35,7 @@ usage() {
     echo "  --name          Container name (Optional, default: $DEFAULT_CONTAINER_NAME)"
     echo "  --eth-if        Ethernet interface (Optional, auto-detected)"
     echo "  --ib-if         InfiniBand interface (Optional, auto-detected)"
+    echo "  -e, --env       Environment variable to pass to container (e.g. -e VAR=val)"
     echo "  --nccl-debug    NCCL debug level (Optional, one of: VERSION, WARN, INFO, TRACE). If no level is provided, defaults to INFO."
     echo "  --apply-mod     Path to directory or zip file containing run.sh to apply before launch (Can be specified multiple times)"
     echo "  --check-config  Check configuration and auto-detection without launching"
@@ -52,6 +53,7 @@ while [[ "$#" -gt 0 ]]; do
         --name) CONTAINER_NAME="$2"; shift ;;
         --eth-if) ETH_IF="$2"; shift ;;
         --ib-if) IB_IF="$2"; shift ;;
+        -e|--env) DOCKER_ARGS="$DOCKER_ARGS -e $2"; shift ;;
         --apply-mod) MOD_PATHS+=("$2"); shift ;;
         --nccl-debug)
             if [[ -n "$2" && "$2" =~ ^(VERSION|WARN|INFO|TRACE)$ ]]; then