Update llama_bootstrap.py

llama_bootstrap.py

@@ -1,49 +1,25 @@
 """
 llama-cpp-python bootstrap — MUST import before gradio or anything else.
+
+2026-07-27: simplified after the great wheel hunt.
+- PyPI = source-only (compiles 20+ min — the old "runs and runs" bug).
+- abetlen CUDA indexes = abandoned (newest cu121 wheel is v0.2.59). Do NOT attempt.
+- abetlen CPU index = current prebuilt wheels. That is the only fast path.
+GPU speed returns via the Level-2 llama-server route (official llama.cpp CUDA
+binaries + MTP flags) — separate spec, not this file.
 """
 
 from __future__ import annotations
 
 import os
-import shutil
 import subprocess
 import sys
 
-CUDA_VARIANTS = (
-    os.environ.get("LLAMA_CPP_CUDA", "cu124"),
-    "cu124",
-    "cu121",
-)
-PIP_BASE = [sys.executable, "-m", "pip"]
-
-
-def has_nvidia_gpu() -> bool:
-    if os.environ.get("LLAMA_CPP_FORCE_CPU", "").strip().lower() in ("1", "true", "yes"):
-        print("[LLAMA] LLAMA_CPP_FORCE_CPU=1 — skipping GPU", flush=True)
-        return False
-    if "CUDA_VISIBLE_DEVICES" in os.environ and os.environ["CUDA_VISIBLE_DEVICES"].strip() == "":
-        print("[LLAMA] CUDA_VISIBLE_DEVICES='' — no GPU visible", flush=True)
-        return False
-    if shutil.which("nvidia-smi") is None:
-        return False
-    try:
-        result = subprocess.run(
-            ["nvidia-smi", "-L"],
-            capture_output=True,
-            text=True,
-            timeout=10,
-        )
-        ok = result.returncode == 0 and bool(result.stdout.strip())
-        if ok:
-            print(f"[LLAMA] GPU detected:\n{result.stdout.strip()}", flush=True)
-        return ok
-    except Exception as e:
-        print(f"[LLAMA] nvidia-smi check failed: {e}", flush=True)
-        return False
+CPU_WHEEL_INDEX = "https://abetlen.github.io/llama-cpp-python/whl/cpu"
 
 
 def _pip(args: list) -> None:
-    cmd = [*PIP_BASE, *args]
+    cmd = [sys.executable, "-m", "pip", *args]
     print(f"[LLAMA] $ {' '.join(cmd)}", flush=True)
     env = os.environ.copy()
     env["PIP_DISABLE_PIP_VERSION_CHECK"] = "1"
@@ -51,67 +27,26 @@ def _pip(args: list) -> None:
     subprocess.check_call(cmd, env=env)
 
 
-def pip_uninstall_llama() -> None:
-    try:
-        _pip(["uninstall", "-y", "llama-cpp-python"])
-    except Exception:
-        pass
-
-
-def pip_install_cpu() -> None:
-    print("[LLAMA] Installing CPU wheel (no CUDA)...", flush=True)
-    _pip(["install", "--no-cache-dir", "llama-cpp-python"])
-
-
-def pip_install_cuda(variant: str) -> None:
-    index = f"https://abetlen.github.io/llama-cpp-python/whl/{variant}"
-    print(f"[LLAMA] Installing CUDA wheel ({variant})...", flush=True)
-    _pip([
-        "install",
-        "--no-cache-dir",
-        "llama-cpp-python",
-        "--extra-index-url",
-        index,
-    ])
-
-
 def try_import_llama():
     import llama_cpp  # noqa: F401
-    from llama_cpp import Llama  # noqa: F401
     return llama_cpp
 
 
 def bootstrap_llama_cpp():
     try:
         mod = try_import_llama()
-        print(f"[LLAMA] Ready (gpu={has_nvidia_gpu()})", flush=True)
-        return mod
+        print("[LLAMA] Ready (prebuilt wheel already installed)", flush=True)
     except ImportError:
-        print("[LLAMA] Not installed — bootstrapping...", flush=True)
-    except (RuntimeError, OSError) as exc:
-        print(f"[LLAMA] Broken install ({exc}) — reinstalling...", flush=True)
-        pip_uninstall_llama()
-
-    if has_nvidia_gpu():
-        last_err = None
-        for variant in CUDA_VARIANTS:
-            try:
-                pip_uninstall_llama()
-                pip_install_cuda(variant)
-                mod = try_import_llama()
-                os.environ.setdefault("DAVIDAU_N_GPU_LAYERS", "-1")
-                print(f"[LLAMA] CUDA ready ({variant}, n_gpu_layers=-1)", flush=True)
-                return mod
-            except Exception as exc:
-                last_err = exc
-                print(f"[LLAMA] CUDA {variant} failed: {exc}", flush=True)
-        print(f"[LLAMA] All CUDA variants failed ({last_err}) — CPU fallback", flush=True)
-        pip_uninstall_llama()
+        print("[LLAMA] Not installed — installing CPU wheel (fast path)...", flush=True)
+        _pip([
+            "install", "--no-cache-dir", "llama-cpp-python==0.3.34",
+            "--extra-index-url", CPU_WHEEL_INDEX,
+        ])
+        mod = try_import_llama()
 
-    pip_install_cpu()
+    # CPU wheel => CPU inference. Be explicit so nothing pretends otherwise.
     os.environ.setdefault("DAVIDAU_N_GPU_LAYERS", "0")
-    mod = try_import_llama()
-    print("[LLAMA] CPU ready (n_gpu_layers=0)", flush=True)
+    print("[LLAMA] CPU mode (n_gpu_layers=0) — GPU speed is the Level-2 llama-server job", flush=True)
     return mod