fix(cookbook): emit HIP_VISIBLE_DEVICES for ROCm serves (#4950)

routes/cookbook_helpers.py

@@ -1078,6 +1078,30 @@ class ServeRequest(BaseModel):
     platform: str | None = None    # "linux", "termux", or "windows"
 
 
+def _serve_gpu_pin_var(env_prefix: str | None, cmd: str | None) -> str:
+    """ROCm/HIP stacks need HIP_VISIBLE_DEVICES; NVIDIA uses CUDA_VISIBLE_DEVICES."""
+    blob = f"{env_prefix or ''} {cmd or ''}"
+    if re.search(r"HIP_VISIBLE_DEVICES|\brocm\b|hipconfig", blob, re.I):
+        return "HIP_VISIBLE_DEVICES"
+    return "CUDA_VISIBLE_DEVICES"
+
+
+def _serve_gpu_export_line(
+    gpus: str | None,
+    env_prefix: str | None,
+    cmd: str | None,
+    *,
+    windows: bool = False,
+) -> str | None:
+    g = (gpus or "").strip()
+    if not g:
+        return None
+    var = _serve_gpu_pin_var(env_prefix, cmd)
+    if windows:
+        return f"$env:{var} = '{g.replace(chr(39), chr(39)+chr(39))}'"
+    return f"export {var}='{_bash_squote(g)}'"
+
+
 def _parse_serve_phase(snapshot: str, task_type: str = "serve") -> dict:
     """Parse a tmux snapshot of a serve task into structured phase info.
 

routes/cookbook_routes.py

@@ -60,7 +60,7 @@
     _user_shell_path_bootstrap, _venv_safe_local_pip_install_cmd,
     _append_pip_install_runner_lines, _pip_install_command_without_break_system_packages,
     _normalize_llama_cpp_python_cache_types,
-    ModelDownloadRequest, ServeRequest,
+    ModelDownloadRequest, ServeRequest, _serve_gpu_export_line,
 )
 
 _HF_TOKEN_STATUS_SNIPPET = (
@@ -2008,7 +2008,9 @@ async def model_serve(request: Request, req: ServeRequest):
             if req.hf_token:
                 ps_lines.append(f"$env:HF_TOKEN = '{_ps_squote(req.hf_token)}'")
             if req.gpus:
-                ps_lines.append(f"$env:CUDA_VISIBLE_DEVICES = '{req.gpus}'")
+                _gpu_line = _serve_gpu_export_line(req.gpus, req.env_prefix, req.cmd, windows=True)
+                if _gpu_line:
+                    ps_lines.append(_gpu_line)
             if req.env_prefix:
                 ps_lines.append(_safe_env_prefix(req.env_prefix))
             # Auto-install ollama if the command uses it
@@ -2082,7 +2084,9 @@ async def model_serve(request: Request, req: ServeRequest):
             if req.hf_token:
                 runner_lines.append(f"export HF_TOKEN='{_bash_squote(req.hf_token)}'")
             if req.gpus:
-                runner_lines.append(f"export CUDA_VISIBLE_DEVICES='{req.gpus}'")
+                _gpu_line = _serve_gpu_export_line(req.gpus, req.env_prefix, req.cmd, windows=False)
+                if _gpu_line:
+                    runner_lines.append(_gpu_line)
             if req.env_prefix:
                 runner_lines.append(_safe_env_prefix(req.env_prefix))
             else:

tests/test_batch4_contrib_fixes.py

@@ -0,0 +1,39 @@
+"""Regression tests for fourth contribution batch."""
+import pytest
+
+from routes.cookbook_helpers import _serve_gpu_pin_var, _serve_gpu_export_line
+from routes.model_routes import _is_discovery_only_provider, _explicit_model_list_timeout
+from services.hwfit.fit import _lookup_bandwidth
+from core.session_manager import _content_for_db_storage
+
+
+def test_hwfit_laptop_4090_bandwidth():
+    bw = _lookup_bandwidth("NVIDIA GeForce RTX 4090 Laptop GPU")
+    assert bw == 504
+
+
+def test_openrouter_discovery_only():
+    assert _is_discovery_only_provider("openrouter") is True
+    assert _is_discovery_only_provider("anthropic") is False
+
+
+def test_ollama_private_lan_timeout():
+    t = _explicit_model_list_timeout("http://192.168.0.5:11434/v1", "auto", None)
+    assert t >= 20.0
+
+
+def test_serve_gpu_hip_when_env_prefix_has_hip():
+    assert _serve_gpu_pin_var("HIP_VISIBLE_DEVICES=1", "llama-server --model x") == "HIP_VISIBLE_DEVICES"
+    line = _serve_gpu_export_line("1", "HIP_VISIBLE_DEVICES=1", "llama-server", windows=False)
+    assert line and "HIP_VISIBLE_DEVICES" in line and "1" in line
+
+
+def test_storage_strips_inline_image_base64():
+    blocks = [
+        {"type": "text", "text": "hi"},
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+    ]
+    out = _content_for_db_storage(blocks)
+    assert isinstance(out, list)
+    assert out[1]["type"] == "text"
+    assert "attachment" in out[1]["text"]