Bleeding edge: vLLM main branch, flashinfer v0.6.7, Gitea fork source

vllm/Dockerfile

@@ -102,7 +102,7 @@ ARG FLASHINFER_ENABLE_AOT=1
 #   - v0.6.7 works with vLLM v0.18.2rc0 (Build #43)
 #   - v0.6.6 works with vLLM v0.19.0 (for Gemma 4 support)
 # ARG FLASHINFER_REF=v0.6.7  # For vLLM v0.18.2rc0
-ARG FLASHINFER_REF=v0.6.6
+ARG FLASHINFER_REF=v0.6.7
 ARG FLASHINFER_BUILD_SUFFIX=cu132
 ENV FLASHINFER_LOCAL_VERSION=${FLASHINFER_BUILD_SUFFIX:-}
 RUN git clone https://github.com/flashinfer-ai/flashinfer.git
@@ -151,10 +151,10 @@ RUN apt-get update && apt-get install -y build-essential cmake gcc && \
 # ==============================================================================
 FROM build-base AS build-vllm
 # vLLM version to build
-ARG VLLM_REF=v0.19.0
+ARG VLLM_REF=main
 # Install ccache for faster compilation
 RUN apt-get update && apt-get install -y ccache
-RUN git clone https://github.com/vllm-project/vllm.git
+RUN git clone https://sweetapi.com/biondizzle/vllm.git
 RUN cd vllm && \
     git checkout ${VLLM_REF} && \
     echo "\n\n========================================" && \

vllm/vllm_shim_module.py

@@ -72,7 +72,7 @@ def download_file(url: str, dest: str):
     for attempt in range(1, MAX_DOWNLOAD_RETRIES + 1):
         try:
             log(f"Downloading {url} -> {dest} (attempt {attempt}/{MAX_DOWNLOAD_RETRIES})")
-            urllib.request.urlretrieve(url, dest, reporthook=_download_progress)
+            urllib.request.urlretrieve(url, dest)
             log(f"Download complete: {dest}")
             return
         except Exception as e:
@@ -88,19 +88,6 @@ def download_file(url: str, dest: str):
                 raise
 
 
-def _download_progress(block_num, block_size, total_size):
-    """Simple download progress callback."""
-    if total_size <= 0:
-        return
-    downloaded = block_num * block_size
-    pct = min(downloaded * 100 // total_size, 100)
-    if pct % 10 == 0 and pct > 0:
-        mb_down = downloaded / (1024 * 1024)
-        mb_total = total_size / (1024 * 1024)
-        sys.stdout.write(f"\r  {pct}% ({mb_down:.0f}/{mb_total:.0f} MB)")
-        sys.stdout.flush()
-
-
 def extract_archive(archive_path: str, dest_dir: str, archive_type: str):
     """Extract archive to dest_dir based on archive_type."""
     log(f"Extracting {archive_path} ({archive_type}) -> {dest_dir}")
@@ -264,12 +251,30 @@ def strip_shim_from_pythonpath():
         log(f"Stripped {SHIM_DIR} from PYTHONPATH (was: {pp!r}, now: {new_pp!r})")
 
 
+def invoked_module_path() -> str:
+    """
+    Derive the dotted module path from this file's location in the shadow package.
+    
+    When invoked via `python -m vllm.entrypoints.openai.api_server`, __name__ is
+    "__main__" — useless for re-invocation. Instead, figure out the module path
+    from the file path relative to the shim root (/opt/vllm-shim).
+    """
+    # e.g. /opt/vllm-shim/vllm/entrypoints/openai/api_server.py
+    filepath = os.path.abspath(__file__)
+    # Strip the shim root + trailing .py, convert / to .
+    rel = os.path.relpath(filepath, SHIM_DIR)
+    # Remove .py extension
+    if rel.endswith(".py"):
+        rel = rel[:-3]
+    return rel.replace(os.sep, ".")
+
+
 def main():
     args = sys.argv[1:]
 
     # Determine which vllm module was actually invoked so we exec the real one
     # (could be vllm.entrypoints.cli.main, vllm.entrypoints.openai.api_server, etc.)
-    invoked_module = __name__  # e.g. "vllm.entrypoints.cli.main" or "vllm.entrypoints.openai.api_server"
+    invoked_module = invoked_module_path()
     log("=" * 50)
     log("vLLM Custom Weights Shim")
     log(f"  Invoked as: python -m {invoked_module} {' '.join(args)}")