vramd 0.2.1__tar.gz → 0.2.2__tar.gz

{vramd-0.2.1/src/vramd.egg-info → vramd-0.2.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vramd
-Version: 0.2.1
+Version: 0.2.2
 Summary: Controlo de admissão de VRAM para inferência generativa em GPUs de consumo
 Author: Maikeu Locatelli
 License-Expression: MIT

{vramd-0.2.1 → vramd-0.2.2}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vramd"
-version = "0.2.1"
+version = "0.2.2"
 description = "Controlo de admissão de VRAM para inferência generativa em GPUs de consumo"
 readme = "README.md"
 authors = [{ name = "Maikeu Locatelli" }]

{vramd-0.2.1 → vramd-0.2.2}/src/vramd/__init__.py

@@ -19,5 +19,5 @@ adivinhar: o número que o admit usa vem do driver, não de uma estimativa.
 
 from __future__ import annotations
 
-__version__ = "0.2.1"
+__version__ = "0.2.2"
 __all__ = ["__version__"]

{vramd-0.2.1 → vramd-0.2.2}/src/vramd/cli.py

@@ -1534,7 +1534,9 @@ def calibrate_cmd(
         elif auto_kwargs:
             console.print(f"[dim]hw-auto ({desc.tool}): {auto_kwargs}[/dim]")
         load_kwargs.update(auto_kwargs)
-    # Explícito vence o automático: o operador pode querer medir outro ponto.
+    # ``calibrate_load_kwargs`` do descriptor (config real de produção do
+    # backend) ganha ao hw-auto; o explícito do CLI ganha a ambos.
+    load_kwargs.update(desc.calibrate_load_kwargs)
     load_kwargs.update(explicit_kwargs)
     if quant:
         load_kwargs.setdefault("sdnq_preset", quant)
@@ -1543,10 +1545,14 @@ def calibrate_cmd(
     if request_json:
         request = json.loads(Path(request_json).read_text(encoding="utf-8"))
     else:
-        request = {
-            "prompt": prompt or "calibration job",
-            "output": output_path or f"/tmp/ums-calibrate-{backend}.bin",
-        }
+        # Default do descriptor (prompt/output com a extensão certa por
+        # backend) — sem isto ``calibrate <backend>`` falha em backends que
+        # exigem inputs (mesh_path/output) ou formatos específicos.
+        request = dict(desc.calibrate_request)
+        request.setdefault("prompt", prompt or "calibration job")
+        if output_path:
+            request["output"] = output_path
+        request.setdefault("output", f"/tmp/vramd-calibrate-{backend}.bin")
 
     pool = SubprocessWorkerPool()
     runner = CalibrationRunner(pool)

{vramd-0.2.1 → vramd-0.2.2}/src/vramd/registry.py

@@ -201,6 +201,15 @@ class BackendDescriptor:
         shape_keys: Subconjunto que força reload quando muda. ``None`` = global.
         vram: Bloco ``vram:`` medido (`vramd calibrate`), quando existe.
         peak_profile: Bloco ``peak_profile:`` (quant, staged, unload_frees_vram…).
+        calibrate_request: Request de geração default para ``vramd calibrate``
+            (ex.: prompt + output com a extensão certa). Sem isto, o calibrador
+            usa ``{}`` e backends que exigem inputs (mesh_path/output) falham
+            com mensagens pouco acionáveis — o default vive no descriptor para
+            que ``vramd calibrate <backend>`` funcione sempre.
+        calibrate_load_kwargs: Kwargs de load default para ``vramd calibrate``
+            (ex.: ``sdnq_preset`` quando o perfil da tool não o resolve).
+            Fundidos com o hw-auto (este ganha) e com os explícitos do CLI
+            (``--load-kwargs``/``--quant`` ganham a este).
     """
 
     name: str
@@ -214,6 +223,8 @@ class BackendDescriptor:
     shape_keys: frozenset[str] | None = None
     vram: Mapping[str, Any] = field(default_factory=dict)
     peak_profile: Mapping[str, Any] = field(default_factory=dict)
+    calibrate_request: Mapping[str, Any] = field(default_factory=dict)
+    calibrate_load_kwargs: Mapping[str, Any] = field(default_factory=dict)
 
     @property
     def unload_frees_vram(self) -> bool:
@@ -316,6 +327,8 @@ def _to_descriptor(entry: Mapping[str, Any], *, source: str) -> BackendDescripto
         shape_keys=frozenset(str(k) for k in shape_keys) if shape_keys else None,
         vram=dict(entry.get("vram") or {}),
         peak_profile=dict(entry.get("peak_profile") or {}),
+        calibrate_request=dict(entry.get("calibrate_request") or {}),
+        calibrate_load_kwargs=dict(entry.get("calibrate_load_kwargs") or {}),
     )
 
 

{vramd-0.2.1 → vramd-0.2.2/src/vramd.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vramd
-Version: 0.2.1
+Version: 0.2.2
 Summary: Controlo de admissão de VRAM para inferência generativa em GPUs de consumo
 Author: Maikeu Locatelli
 License-Expression: MIT

{vramd-0.2.1 → vramd-0.2.2}/tests/test_calibrate_cli.py

@@ -115,6 +115,70 @@ class TestCalibrateCommand:
         assert result.exit_code == 0
         assert patched["spec"].request["prompt"] == "uma pedra"
 
+    def test_descriptor_calibrate_defaults_are_used(self, runner, patched, tmp_path, monkeypatch):
+        """`vramd calibrate <backend>` funciona sem flags quando o descriptor
+        declara ``calibrate_request``/``calibrate_load_kwargs`` — o utilizador
+        comum não precisa de saber que o backend exige mesh_path/output/quant."""
+        yaml_path = tmp_path / "backends.yaml"
+        yaml_path.write_text(
+            yaml.safe_dump(
+                {
+                    "version": 2,
+                    "backends": [
+                        {
+                            "name": "text3d",
+                            "adapter": "x",
+                            "vram_mib": 4000,
+                            "tool": "text3d",
+                            "calibrate_request": {"mesh_path": "/tmp/m.glb", "output": "/tmp/out.glb"},
+                            "calibrate_load_kwargs": {"sdnq_preset": "sdnq-int4"},
+                        }
+                    ],
+                }
+            ),
+            encoding="utf-8",
+        )
+        monkeypatch.setenv("VRAMD_BACKENDS_FILE", str(yaml_path))
+        result = runner.invoke(cli_mod.cli, ["calibrate", "text3d", "--no-compare"])
+        assert result.exit_code == 0
+        spec = patched["spec"]
+        assert spec.request["mesh_path"] == "/tmp/m.glb"
+        assert spec.request["output"] == "/tmp/out.glb"
+        assert spec.load_kwargs["sdnq_preset"] == "sdnq-int4"
+
+    def test_explicit_flags_override_descriptor_defaults(self, runner, patched, tmp_path, monkeypatch):
+        """--request-json/--load-kwargs do operador ganham ao descriptor."""
+        yaml_path = tmp_path / "backends.yaml"
+        yaml_path.write_text(
+            yaml.safe_dump(
+                {
+                    "version": 2,
+                    "backends": [
+                        {
+                            "name": "text3d",
+                            "adapter": "x",
+                            "vram_mib": 4000,
+                            "tool": "text3d",
+                            "calibrate_request": {"mesh_path": "/a.glb", "output": "/a-out.glb"},
+                            "calibrate_load_kwargs": {"sdnq_preset": "sdnq-int4"},
+                        }
+                    ],
+                }
+            ),
+            encoding="utf-8",
+        )
+        monkeypatch.setenv("VRAMD_BACKENDS_FILE", str(yaml_path))
+        job = tmp_path / "job.json"
+        job.write_text(json.dumps({"mesh_path": "/b.glb", "output": "/b-out.glb"}), encoding="utf-8")
+        result = runner.invoke(
+            cli_mod.cli,
+            ["calibrate", "text3d", "--no-compare", "--request-json", str(job), "--load-kwarg", "sdnq_preset=none"],
+        )
+        assert result.exit_code == 0
+        spec = patched["spec"]
+        assert spec.request["mesh_path"] == "/b.glb"
+        assert spec.load_kwargs["sdnq_preset"] is None  # "none" é coergido a None
+
     def test_repeats_and_cycles_flow_through(self, runner, patched):
         runner.invoke(cli_mod.cli, ["calibrate", "text3d", "--no-compare", "--repeats", "5", "--cycles", "2"])
         assert patched["spec"].repeats == 5