xpu (#4969)

fastdeploy/model_executor/utils.py

@@ -353,6 +353,13 @@ def h2d_copy(dst, src, blocking=True):
 def v1_loader_support(fd_config):
     _v1_no_support_archs = ["Qwen2VLForConditionalGeneration"]
 
+    def _get_unsupported_quant():
+        if current_platform.is_cuda():
+            return {"w4a8", "w4afp8", "wint2"}
+        elif current_platform.is_xpu():
+            return {"w4a8", "w8a8"}
+        return set()
+
     def _err_msg(msg: str) -> str:
         logger.info(msg + "; fallback to the v0 loader for model loading.")
 
@@ -375,7 +382,7 @@ def v1_loader_support(fd_config):
         else:
             moe_quant_type = fd_config.quant_config.name()
             dense_quant_type = fd_config.quant_config.name()
-        unsupported_quant = {"w4a8", "w4afp8", "wint2"}
+        unsupported_quant = _get_unsupported_quant()
 
         if unsupported_quant & {moe_quant_type, dense_quant_type}:
             _err_msg("v1 loader currently does not support w4a8/w4afp8/win2 quantization")

requirements_dcu.txt

@@ -27,6 +27,7 @@ moviepy
 use-triton-in-paddle
 crcmod
 fastsafetensors==0.1.14
+safetensors==0.7.0rc0
 msgpack
 gunicorn
 opentelemetry-api>=1.24.0

requirements_metaxgpu.txt

@@ -29,6 +29,7 @@ triton
 use-triton-in-paddle
 crcmod
 fastsafetensors==0.1.14
+safetensors==0.7.0rc0
 msgpack
 gunicorn
 modelscope

scripts/run_ci_xpu.sh

@@ -81,8 +81,7 @@ python -m fastdeploy.entrypoints.openai.api_server \
     --num-gpu-blocks-override 16384 \
     --max-model-len 32768 \
     --max-num-seqs 128 \
-    --quantization wint4  \
-    --load-choices default  > server.log 2>&1 &
+    --quantization wint4 > server.log 2>&1 &
 
 sleep 60
 # 探活
@@ -157,8 +156,7 @@ python -m fastdeploy.entrypoints.openai.api_server \
     --num-gpu-blocks-override 16384 \
     --max-model-len 32768 \
     --max-num-seqs 64 \
-    --quantization "W4A8" \
-    --load-choices default  > server.log 2>&1 &
+    --quantization "W4A8"   > server.log 2>&1 &
 
 sleep 60
 # 探活
@@ -236,8 +234,7 @@ python -m fastdeploy.entrypoints.openai.api_server \
     --enable-mm \
     --mm-processor-kwargs '{"video_max_frames": 30}' \
     --limit-mm-per-prompt '{"image": 10, "video": 3}' \
-    --reasoning-parser ernie-45-vl \
-    --load-choices default > server.log 2>&1 &
+    --reasoning-parser ernie-45-vl > server.log 2>&1 &
 
 sleep 60
 # 探活