Add --offload-strategy none to disable CPU offloading and VRAM auto-detection

- Add 'none' to --offload-strategy choices in cli.py
- In cuda.py backend:
  - _get_vram_percentages_for_strategy() returns None for 'none' strategy
  - _get_vram_percentages_for_gpu() skips VRAM detection for 'none'
  - load_model() loads directly on GPU without max_memory constraints
- Add startup status message in main.py for --offload-strategy none

codai/backends/cuda.py

@@ -185,6 +185,9 @@ class NvidiaBackend(ModelBackend):
     
     def _get_vram_percentages_for_strategy(self, strategy: str, is_moe: bool, total_vram_gb: float) -> list:
         """Get VRAM percentage steps based on offload strategy."""
+        if strategy == "none":
+            print(f"  Offload strategy 'none': disabling CPU offload and VRAM auto-detection")
+            return None  # Signal to skip offloading entirely
         if strategy == "conservative":
             print(f"  Using conservative offload strategy")
             if is_moe:
@@ -221,9 +224,15 @@ class NvidiaBackend(ModelBackend):
                     return [0.93, 0.85, 0.75, 0.65, 0.50, 0.35, 0.20, 0.0]
     
     def _get_vram_percentages_for_gpu(self, model_name: str = "", strategy: str = "auto", max_gpu_percent: float = None) -> list:
-        """Get VRAM percentage steps based on GPU memory size."""
+        """Get VRAM percentage steps based on GPU memory size.
+        
+        Returns None when strategy is 'none' (no offloading).
+        """
         import torch
         
+        if strategy == "none":
+            return None  # Signal to skip offloading entirely
+        
         if not torch.cuda.is_available():
             return [0.0]
         
@@ -397,6 +406,25 @@ class NvidiaBackend(ModelBackend):
         
         model = None
         vram_percentages = self._get_vram_percentages_for_gpu(model_name, offload_strategy, max_gpu_percent)
+        
+        # --offload-strategy none: load directly on GPU without offloading or VRAM limits
+        if vram_percentages is None:
+            cuda_device = self._derive_cuda_device()
+            print(f"\nOffload strategy 'none': loading model directly on {cuda_device} (no CPU offload, no VRAM limits)")
+            load_kwargs['device_map'] = cuda_device
+            load_kwargs['low_cpu_mem_usage'] = True
+            load_kwargs['torch_dtype'] = "auto"
+            # Remove dtype set earlier since torch_dtype=auto takes precedence
+            load_kwargs.pop('dtype', None)
+            
+            try:
+                model = AutoModelForCausalLM.from_pretrained(model_name, **load_kwargs)
+            except Exception as e:
+                raise RuntimeError(
+                    f"--offload-strategy none: Failed to load model entirely on GPU ({cuda_device}). "
+                    f"The model may be too large for available VRAM. Error: {e}"
+                )
+        else:
             first_vram_pct = vram_percentages[0] if vram_percentages else 0.93
             
             for vram_pct in vram_percentages:

codai/cli.py

@@ -117,9 +117,9 @@ def parse_args():
     parser.add_argument(
         "--offload-strategy",
         type=str,
-        choices=["auto", "conservative", "balanced", "aggressive", "sequential"],
+        choices=["auto", "conservative", "balanced", "aggressive", "sequential", "none"],
         default="auto",
-        help="Offload strategy for NVIDIA backend (default: auto)",
+        help="Offload strategy for NVIDIA backend (default: auto). Use 'none' to disable CPU offloading and VRAM auto-detection entirely.",
     )
     parser.add_argument(
         "--max-gpu-percent",

codai/main.py

@@ -187,6 +187,11 @@ def main():
     if grammar_guided_gen:
         print("Grammar-guided generation enabled (--grammar-guided-gen)")
     
+    # Print --offload-strategy none status
+    if args.offload_strategy == "none":
+        print("Offload strategy 'none': CPU offloading and VRAM auto-detection disabled")
+        print("  Model will be loaded directly on GPU without memory limits")
+    
     # Print --no-ram mode status
     if args.no_ram:
         print("No-RAM mode enabled (--no-ram): maximizing VRAM usage, no CPU RAM spilling")