fix: pin Gradio NanoVLLM runtime dependencies

Avoid the torch 2.11 CUDA 13 stack that breaks flash-attn builds on Spaces by pinning a more compatible torch/torchaudio/triton set and installing a pinned flash-attn wheel without dependency upgrades.

Made-with: Cursor

app.py

@@ -82,7 +82,7 @@ def _ensure_nanovllm_runtime() -> None:
     try:
         import flash_attn  # noqa: F401
     except ImportError:
-        logger.info("Installing flash-attn at runtime ...")
+        logger.info("Installing pinned flash-attn at runtime ...")
         subprocess.check_call(
             [
                 sys.executable,
@@ -90,7 +90,8 @@ def _ensure_nanovllm_runtime() -> None:
                 "pip",
                 "install",
                 "--no-build-isolation",
-                "flash-attn",
+                "--no-deps",
+                "flash-attn==2.6.3",
             ]
         )
 

requirements.txt

@@ -2,9 +2,9 @@ gradio==6.0.0
 huggingface-hub
 funasr
 numpy>=1.21.0
-torch>=2.5.0,!=2.6.*
-torchaudio
-triton>=3.0.0
+torch==2.5.1
+torchaudio==2.5.1
+triton==3.1.0
 transformers>=4.51.0
 xxhash
 tqdm