fix: update CUDA to 12.4.1 for Blackwell GPU support (#251)

* fix: update CUDA to 12.4.1 for Blackwell GPU support

- Update Dockerfile base image from CUDA 12.1.0 to 12.4.1
- Update ldconfig path to cuda-12.4
- Update FlashInfer installation to use flashinfer-python package
- Add NVIDIA B200 (Blackwell) to supported gpuIds in hub.json

This fixes the "imagePullAsync: failed to get self-hosted image registry auth"
error when deploying on Blackwell GPUs (RTX PRO 6000, B200) by aligning
the Docker image CUDA version with the allowedCudaVersions in hub.json.

Fixes: DR-1118

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* revert: remove NVIDIA B200 from default gpuIds

The gpuIds in hub.json controls default GPU selection for deployments,
not GPU compatibility. The CUDA 12.4 upgrade is sufficient to enable
Blackwell GPU support.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

* fix: remove FlashInfer to avoid JIT compilation errors

FlashInfer requires nvcc to JIT-compile CUDA kernels at runtime for
new GPU architectures (like Blackwell SM 10.0). Since we use the CUDA
base image without the toolkit, nvcc is not available.

vLLM will use its built-in fallback sampling methods instead.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.5 <noreply@anthropic.com>

.runpod/hub.json

@@ -929,6 +929,16 @@
           "advanced": true
         }
       },
+      {
+        "key": "ENABLE_EXPERT_PARALLEL",
+        "input": {
+          "name": "Enable Expert Parallel",
+          "type": "boolean",
+          "description": "Enable Expert Parallel for MoE models",
+          "default": false,
+          "advanced": true
+        }
+      },
       {
         "key": "MODEL_REVISION",
         "input": {

Dockerfile

@@ -1,9 +1,9 @@
-FROM nvidia/cuda:12.1.0-base-ubuntu22.04 
+FROM nvidia/cuda:12.4.1-base-ubuntu22.04 
 
 RUN apt-get update -y \
     && apt-get install -y python3-pip
 
-RUN ldconfig /usr/local/cuda-12.1/compat/
+RUN ldconfig /usr/local/cuda-12.4/compat/
 
 # Install Python dependencies
 COPY builder/requirements.txt /requirements.txt
@@ -11,9 +11,8 @@ RUN --mount=type=cache,target=/root/.cache/pip \
     python3 -m pip install --upgrade pip && \
     python3 -m pip install --upgrade -r /requirements.txt
 
-# Install vLLM (switching back to pip installs since issues that required building fork are fixed and space optimization is not as important since caching) and FlashInfer 
-RUN python3 -m pip install vllm==0.11.0 && \
-    python3 -m pip install flashinfer -i https://flashinfer.ai/whl/cu121/torch2.3
+# Install vLLM
+RUN python3 -m pip install vllm==0.11.0
 
 # Setup for Option 2: Building the Image with the Model included
 ARG MODEL_NAME=""

builder/requirements.txt

@@ -1,7 +1,7 @@
 ray
 pandas
 pyarrow
-runpod~=1.7.7
+runpod>=1.8,<2.0
 huggingface-hub
 packaging
 typing-extensions>=4.8.0

docs/configuration.md

@@ -85,6 +85,7 @@ Complete guide to all environment variables and configuration options for worker
 | `ENFORCE_EAGER`                | False   | `bool`          | Always use eager-mode PyTorch. If False(`0`), will use eager mode and CUDA graph in hybrid for maximal performance and flexibility. |
 | `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  |
 
 ## Tokenizer Settings
 

src/engine_args.py

@@ -80,6 +80,7 @@ DEFAULT_ARGS = {
     "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,