feat: swap to fine-tuned Q4_K_M model (4.68 GB)

- Q4_K_M quantization complete via llama.cpp
- model_inference.py now uses lfed-qwen2.5-coder-7b-sql-Q4_K_M.gguf
- 4.68 GB vs 15.2 GB F16 — fits Zero GPU comfortably

modal_train/export_gguf.py

@@ -111,15 +111,15 @@ def convert_to_gguf():
         "--outtype", "f16",
     ], check=True)
 
-    # Quantize using llama-quantize (built-in to llama-cpp-python)
+    # Quantize using llama-cpp-python's built-in quantizer
     print(f"   Quantizing to {GGUF_QUANT}...")
-    quantize_bin = shutil.which("llama-quantize") or shutil.which("quantize")
-    if quantize_bin:
-        subprocess.run([quantize_bin, str(f16_path), str(out_path), GGUF_QUANT], check=True)
-    else:
-        # Fallback: use the FP16 GGUF directly
-        print("   ⚠️  llama-quantize not found, using FP16 GGUF")
-        out_path = f16_path
+    from llama_cpp import llama_quantize
+    llama_quantize(
+        input_path=str(f16_path),
+        output_path=str(out_path),
+        output_type=GGUF_QUANT.lower(),
+    )
+    print(f"   Quantized: {out_path}")
 
     print(f"✅ GGUF saved to {out_path}")
     return out_path

modal_train/modal_app.py

@@ -18,7 +18,7 @@ import modal
 
 # ── Modal app ──────────────────────────────────────────────────────────
 
-app = modal.App("kasualdad-lfed-train")
+app = modal.App("kasualdad-lfed-train")  # force redeploy v3
 
 # ── Image with all dependencies ────────────────────────────────────────
 

model_inference.py

@@ -99,7 +99,7 @@ LOCAL_MODEL_PATH = "/tmp/lfed-models/qwen/Qwen2.5-Coder-7B-Instruct.Q4_K_M.gguf"
 
 # Fallback: download from HF Hub (base model, pre-fine-tune)
 HF_REPO_ID = "build-small-hackathon/lfed-qwen2.5-coder-7b-sql-gguf"
-HF_MODEL_FILE = "model-f16.gguf"
+HF_MODEL_FILE = "lfed-qwen2.5-coder-7b-sql-Q4_K_M.gguf"
 
 # Post-Phase-7: swap HF_REPO_ID / HF_MODEL_FILE to the fine-tuned GGUF
 # after Modal training completes and the model is pushed to HF Hub.