[BugFix] fix flashinfer-cutedsl moe nvfp4 (#7120)

* fix nvfp4 * fix * add document * fix nvfp4 * support eb5 * support bka * support eb5 * support xpu * fix * fix * add import cutedsl * fix * fix * fix test * fix H卡 * update document * fix * update document * update document * fix
2026-04-23 00:17:25 +08:00 · 2026-04-03 15:43:19 +08:00
parent 095a11d932
commit 5f612a348d
8 changed files with 317 additions and 90 deletions
@@ -18,7 +18,38 @@ Based on [FlashInfer](https://github.com/flashinfer-ai/flashinfer), Fastdeploy s
 Please ensure that FastDeploy is installed with NVIDIA GPU support.
 Follow the official guide to set up the base environment: [Fastdeploy NVIDIA GPU Environment Installation Guide](https://paddlepaddle.github.io/FastDeploy/get_started/installation/nvidia_gpu/).

-### Running Inference Service
+### FlashInfer-cutedsl backend
+
+#### PaddlePaddle Compatibility Patches for FlashInfer
+
+Due to compatibility issues between FlashInfer and PaddlePaddle, you need to apply the following patches in `miniconda/envs/<your_env>/lib/python3.10/site-packages/`:
+
+1. **nvidia_cutlass_dsl/python_packages/cutlass/torch.py**
+
+   Replace `torch.device` with `"torch.device"` (as a string to avoid conflicts).
+
+2. **flashinfer/utils.py**
+
+  Modify the `get_compute_capability` function:
+  ```bash
+  @functools.cache
+  def get_compute_capability(device: torch.device) -> Tuple[int, int]:
+      return torch.cuda.get_device_capability(device)
+      if device.type != "cuda":
+          raise ValueError("device must be a cuda device")
+      return torch.cuda.get_device_capability(device.index)
+  ```
+
+3. **flashinfer/cute_dsl/blockscaled_gemm.py**
+
+  Replace `cutlass_torch.current_stream()` with:
+  ```bash
+  cuda.CUstream(torch.cuda.current_stream().stream_base.raw_stream)
+  ```
+
+#### Running Inference Service
+
+flashinfer-cutlass backend:
 ```bash
 python -m fastdeploy.entrypoints.openai.api_server \
    --model nv-community/Qwen3-30B-A3B-FP4 \
@@ -31,6 +62,26 @@ python -m fastdeploy.entrypoints.openai.api_server \
    --max-num-seqs 128
 ```

+flashinfer-cutedsl backend:
+```bash
+python -m fastdeploy.entrypoints.openai.multi_api_server \
+       --ports "9811,9812,9813,9814" \
+       --num-servers 4 \
+       --model ERNIE-4.5-21B-A3B-FP4 \
+       --disable-custom-all-reduce \
+       --tensor-parallel-size 1 \
+       --data-parallel-size 4 \
+       --no-enable-prefix-caching \
+       --max-model-len 65536 \
+       --enable-expert-parallel \
+       --num-gpu-blocks-override 8192 \
+       --max-num-seqs 4 \
+       --gpu-memory-utilization 0.9 \
+       --max-num-batched-tokens 512 \
+       --ep-prefill-use-worst-num-tokens \
+       --graph-optimization-config '{"use_cudagraph":false}'
+```
+
 ### API Access
 Make service requests using the following command

@@ -43,6 +94,15 @@ curl -X POST "http://0.0.0.0:8180/v1/chat/completions" \
  ]
 }'
 ```
+```shell
+curl -X POST "http://0.0.0.0:9811/v1/chat/completions" \
+-H "Content-Type: application/json" \
+-d '{
+  "messages": [
+    {"role": "user", "content": "把李白的静夜思改写为现代诗"}
+  ]
+}'
+```

 FastDeploy service interface is compatible with OpenAI protocol. You can make service requests using the following Python code.

@@ -64,4 +124,4 @@ for chunk in response:
    if chunk.choices[0].delta:
        print(chunk.choices[0].delta.content, end='')
 print('\n')
-```.
+```