Fixed gpu detection for cuda rocm etc using env vars

Signed-off-by: Brian <[email protected]>

ramalama/model.py

@@ -98,12 +98,14 @@ def _image(self, args):
         if args.image != default_image():
             return args.image
 
-        gpu_type, _ = get_gpu()
-        if gpu_type == "HIP_VISIBLE_DEVICES":
+        if os.getenv("HIP_VISIBLE_DEVICES"):
             return "quay.io/ramalama/rocm:latest"
 
-        if gpu_type == "ASAHI_VISIBLE_DEVICES":
+        if os.getenv("ASAHI_VISIBLE_DEVICES"):
             return "quay.io/ramalama/asahi:latest"
+
+        if os.getenv("CUDA_VISIBLE_DEVICES"):
+            return "docker.io/brianmahabir/rama-cuda:v1"
 
         return args.image
 
@@ -143,9 +145,15 @@ def setup_container(self, args):
         if os.path.exists("/dev/kfd"):
             conman_args += ["--device", "/dev/kfd"]
 
-        gpu_type, gpu_num = get_gpu()
-        if gpu_type == "HIP_VISIBLE_DEVICES" or gpu_type == "ASAHI_VISIBLE_DEVICES":
-            conman_args += ["-e", f"{gpu_type}={gpu_num}"]
+        for var in ["HIP_VISIBLE_DEVICES", "ASAHI_VISIBLE_DEVICES", "CUDA_VISIBLE_DEVICES"]:
+            value = os.getenv(var)
+            if value:
+                if var == "CUDA_VISIBLE_DEVICES":
+                    # Special handling for CUDA (e.g., using '--gpus all')
+                    conman_args += ["--gpus", "all"]
+                else:
+                    # For HIP and ASAHI, we directly add the environment variable with its value
+                    conman_args += ["-e", f"{var}={value}"]
         return conman_args
 
     def run_container(self, args, shortnames):
@@ -190,14 +198,14 @@ def cleanup():
         return True
 
     def gpu_args(self):
+        gpu_type, gpu_num = get_gpu()
         gpu_args = []
         if sys.platform == "darwin":
             # llama.cpp will default to the Metal backend on macOS, so we don't need
             # any additional arguments.
             pass
-        elif sys.platform == "linux" and (
-            os.getenv("HIP_VISIBLE_DEVICES") or os.getenv("ASAHI_VISIBLE_DEVICES") or os.getenv("CUDA_VISIBLE_DEVICES")
-        ):
+        elif sys.platform == "linux" and gpu_type is not None:
+            os.environ[gpu_type] = gpu_num
             gpu_args = ["-ngl", "99"]
         else:
             print("GPU offload was requested but is not available on this system")
@@ -382,31 +390,82 @@ def check_valid_model_path(self, relative_target_path, model_path):
         return os.path.exists(model_path) and os.readlink(model_path) == relative_target_path
 
 
+def initialize_gpu_template():
+    """Initialize a GPU template."""
+    return {
+        "gpus": []  # List of GPU entries, each with index, memory, and environment variables
+    }
+
+def add_gpu_entry(template, gpu_arch, index, memory_mib, env_var):
+    """Add a GPU entry to the template if it has valid memory."""
+    if memory_mib > 1024:  # Only consider GPUs with > 1 GiB of memory
+        template["gpus"].append({
+            "gpu_arch": gpu_arch,
+            "index": index,
+            "memory_mib": memory_mib,
+            "env_var": env_var
+        })
+
+def get_amdgpu(template):
+    """Detect AMD GPUs and append valid entries to the template."""
+    for i, fp in enumerate(sorted(glob.glob('/sys/bus/pci/devices/*/mem_info_vram_total'))):
+        try:
+            with open(fp, 'r') as file:
+                memory_bytes = int(file.read())
+                memory_mib = memory_bytes / (1024 * 1024)  # Convert bytes to MiB
+                add_gpu_entry(template, "AMD", str(i), memory_mib, "HIP_VISIBLE_DEVICES")
+        except Exception as ex:
+            print(f"Error reading AMD GPU memory info: {ex}")
+
+def get_nvgpu(template):
+    """Detect NVIDIA GPUs and append valid entries to the template."""
+    try:
+        command = ['nvidia-smi', '--query-gpu=index,memory.total', '--format=csv,noheader,nounits']
+        output = run_cmd(command).stdout.decode("utf-8")
+        for line in output.strip().split('\n'):
+            try:
+                index, memory_mib = line.split(',')
+                memory_mib = int(memory_mib)
+                add_gpu_entry(template, "NVIDIA", index.strip(), memory_mib, "CUDA_VISIBLE_DEVICES")
+            except ValueError as ex:
+                print(f"Error parsing NVIDIA GPU info: {ex}")
+    except FileNotFoundError:
+        print("The 'nvidia-smi' command was not found.")
+    except Exception as ex:
+        print(f"An unexpected error occurred while checking for NVIDIA GPUs: {ex}")
+
 def get_gpu():
-    i = 0
-    gpu_num = 0
-    gpu_bytes = 0
-    for fp in sorted(glob.glob('/sys/bus/pci/devices/*/mem_info_vram_total')):
-        with open(fp, 'r') as file:
-            content = int(file.read())
-            if content > 1073741824 and content > gpu_bytes:
-                gpu_bytes = content
-                gpu_num = i
+    """
+    Detects and selects a GPU with at least 1 GiB of memory.
+    Uses a centralized template to handle multiple GPU types.
+    
+    Returns:
+        tuple: Environment variable name and GPU index (as a string), or (None, None) if no suitable GPU is found.
+    """
+    # Check if system is running Asahi Linux (Apple Silicon)
+    if os.path.exists('/etc/os-release'):
+        try:
+            with open('/etc/os-release', 'r') as file:
+                if "asahi" in file.read().lower():
+                    return "ASAHI_VISIBLE_DEVICES", "1"
+        except Exception as ex:
+            print(f"Error reading OS release file: {ex}")
 
-        i += 1
+    # Initialize the GPU template
+    gpu_template = initialize_gpu_template()
 
-    if gpu_bytes:  # this is the ROCm/AMD case
-        return "HIP_VISIBLE_DEVICES", gpu_num
+    # Detect GPUs from different architectures
+    get_amdgpu(gpu_template)
+    get_nvgpu(gpu_template)
 
-    if os.path.exists('/etc/os-release'):
-        with open('/etc/os-release', 'r') as file:
-            content = file.read()
-            if "asahi" in content.lower():
-                return "ASAHI_VISIBLE_DEVICES", 1
+    # Sort GPUs by memory (descending order) and return the best one
+    if gpu_template["gpus"]:
+        best_gpu = sorted(gpu_template["gpus"], key=lambda x: x["memory_mib"], reverse=True)[0]
+        return best_gpu["env_var"], best_gpu["index"]
 
+    # No suitable GPU found
     return None, None
 
-
 def dry_run(args):
     for arg in args:
         if not arg: