Update to new version of base image

Dockerfile

@@ -1,19 +1,15 @@
-FROM git.modelhub.org.cn:9443/enginex-iluvatar/bi100-3.2.1-x86-ubuntu20.04-py3.10-poc-llm-infer:20250731115755
+FROM git.modelhub.org.cn:9443/enginex-iluvatar/bi100-3.2.3-x86-ubuntu20.04-py3.10-poc-llm-infer:v1.2.3
 
 RUN pip install --no-cache-dir triton==2.1.0
 
 COPY pkgs/triton /usr/local/corex/lib64/python3/dist-packages/triton
 COPY pkgs/triton-2.1.0+corex.4.1.2.dist-info /usr/local/corex/lib64/python3/dist-packages/triton-2.1.0+corex.4.1.2.dist-info
-COPY pkgs/xformers-0.0.22+corex.4.1.2.dist-info /usr/local/corex/lib64/python3/dist-packages/xformers-0.0.22+corex.4.1.2.dist-info
-COPY pkgs/xformers /usr/local/corex/lib64/python3/dist-packages/xformers
 
-COPY paged_attn.py /usr/local/lib/python3.10/site-packages/vllm/attention/ops/paged_attn.py
+COPY paged_attn.py /usr/local/corex/lib64/python3/dist-packages/vllm/attention/ops/paged_attn.py
-COPY __init__.py /usr/local/lib/python3.10/site-packages/vllm/triton_utils/__init__.py
+COPY __init__.py /usr/local/corex/lib64/python3/dist-packages/vllm/triton_utils/__init__.py
-COPY prefix_prefill.py /usr/local/lib/python3.10/site-packages/vllm/attention/ops/prefix_prefill.py
 
 RUN mkdir /workspace
 WORKDIR /workspace/
 
 COPY ./launch_service /workspace/launch_service
 
-ENTRYPOINT ["./launch_service"]

paged_attn.py

@@ -4,6 +4,7 @@ from typing import List, Optional, Tuple
 import torch
 
 from vllm import _custom_ops as ops
+
 from vllm.attention.ops.prefix_prefill import context_attention_fwd
 
 # Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.

vllm/_custom_ops.py

@@ -1102,4 +1102,4 @@ for k, v in names_and_values.items():
         names_and_values_to_update[k] = hint_on_error(v)
 
 names_and_values.update(names_and_values_to_update)
-del names_and_values_to_update, names_and_values, v, k, fn_type
+del names_and_values_to_update, names_and_values, v, k, fn_type

vllm/assets/image.py

@@ -27,4 +27,4 @@ class ImageAsset:
         """
         image_path = get_vllm_public_assets(filename=f"{self.name}.pt",
                                             s3_prefix=VLM_IMAGES_DIR)
-        return torch.load(image_path)
+        return torch.load(image_path, weights_only=True)

vllm/attention/ops/paged_attn.py

@@ -4,7 +4,10 @@ from typing import List, Optional, Tuple
 import torch
 
 from vllm import _custom_ops as ops
-from vllm.attention.ops.prefix_prefill import context_attention_fwd
+from vllm.triton_utils import HAS_TRITON
+
+if HAS_TRITON:
+    from vllm.attention.ops.prefix_prefill import context_attention_fwd
 
 # Should be the same as PARTITION_SIZE in `paged_attention_v2_launcher`.
 _PARTITION_SIZE = 512

vllm/attention/ops/prefix_prefill.py

@@ -808,8 +808,6 @@ if triton.__version__ >= "2.1.0":
             )
             return
 
-        import time
-        ts_beg = time.time()
         _fwd_kernel[grid](
             q,
             k,
@@ -860,6 +858,4 @@ if triton.__version__ >= "2.1.0":
             num_warps=NUM_WARPS,
             num_stages=1,
         )
-        elapsed = time.time() - ts_beg
-        #print(f'{elapsed}: {BLOCK=}, {Lk=}, {Lk_padded=}, {BLOCK=}, {sliding_window=}, {NUM_WARPS=}')
         return

vllm/attention/selector.py

@@ -204,9 +204,6 @@ def which_attn_to_use(
         if selected_backend != _Backend.PALLAS:
             logger.info("Cannot use %s backend on TPU.", selected_backend)
         return _Backend.PALLAS
-    
-    if selected_backend == _Backend.FLASH_ATTN:
-        print("selected_backend == _Backend.FLASH_ATTN")
 
     if is_hip():
         # AMD GPUs.

vllm/distributed/device_communicators/shm_broadcast.py

@@ -217,7 +217,8 @@ class MessageQueue:
             remote_subscribe_port = get_open_port()
             if is_valid_ipv6_address(connect_ip):
                 self.remote_socket.setsockopt(IPV6, 1)
-            socket_addr = f"tcp://*:{remote_subscribe_port}"
+                connect_ip = f"[{connect_ip}]"
+            socket_addr = f"tcp://{connect_ip}:{remote_subscribe_port}"
             self.remote_socket.bind(socket_addr)
 
         else: