vramd 0.2.2__tar.gz → 0.2.4__tar.gz

vramd-0.2.4/PKG-INFO

@@ -0,0 +1,245 @@
+Metadata-Version: 2.4
+Name: vramd
+Version: 0.2.4
+Summary: VRAM admission control for generative inference on consumer GPUs
+Author: Maikeu Locatelli
+License-Expression: MIT
+Project-URL: Homepage, https://github.com/maikramer/vramd
+Project-URL: Repository, https://github.com/maikramer/vramd
+Project-URL: Issues, https://github.com/maikramer/vramd/issues
+Project-URL: Changelog, https://github.com/maikramer/vramd/blob/main/CHANGELOG.md
+Keywords: gpu,vram,inference,scheduler,admission-control,diffusion
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: System :: Distributed Computing
+Requires-Python: >=3.11
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: click>=8.1.0
+Requires-Dist: rich>=13.0.0
+Requires-Dist: rich-click>=1.8.0
+Requires-Dist: PyYAML>=6.0
+Requires-Dist: psutil>=5.9.0
+Requires-Dist: nvidia-ml-py>=12.535.0
+Provides-Extra: dev
+Requires-Dist: pytest>=7.4.0; extra == "dev"
+Requires-Dist: pytest-cov>=4.0.0; extra == "dev"
+Requires-Dist: ruff==0.16.2; extra == "dev"
+Dynamic: license-file
+
+# vramd
+
+[![CI](https://github.com/maikramer/vramd/actions/workflows/ci.yml/badge.svg)](https://github.com/maikramer/vramd/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/vramd.svg)](https://pypi.org/project/vramd/)
+[![Python](https://img.shields.io/pypi/pyversions/vramd.svg)](https://pypi.org/project/vramd/)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+
+**VRAM admission control for generative inference on consumer GPUs.**
+
+One process holds the GPU and decides who gets in. It admits by the **real
+peak** — weights + activation + margin, not just weights — queues with
+priority and affinity, evicts by weight+LRU, and runs each model in its own
+process and venv.
+
+Built for inference that lasts **seconds to minutes** on a card that can't fit
+everything. It's not an LLM server: it doesn't optimize token throughput, it
+optimizes *fitting*.
+
+```bash
+pip install vramd          # 9 MB — the supervisor doesn't import torch
+vramd start &
+vramd submit my-model --prompt "…" --wait
+```
+
+## The problem
+
+You have 6 GB of VRAM and five models that, together, ask for 40. Each runs
+fine on its own. Together, the second job starts in the middle of the first
+and both die with `CUDA out of memory` — after the weights were already
+loaded.
+
+The usual solutions assume something that isn't true here: that the model fits
+(vLLM, TGI), that the environment is homogeneous (Ray Serve, Triton), or that
+the unit of work is a token rather than a two-minute job.
+
+## What vramd does differently
+
+**Admits by the peak, not the weights.** The question isn't "does the model
+fit" — it's "does the inference peak fit". It's the difference between
+refusing in 0.2 s and dying at 80% of the job with the weights already
+loaded.
+
+**Each model in its own venv.** A backend is a process with its own
+interpreter. Models with incompatible dependencies — torch 2.x vs 2.y,
+different CUDA wheels — coexist without seeing each other.
+
+**Affinity in the queue.** If the head of the queue needs a cold model and a
+job further back needs one already in VRAM, the scheduler skips the head (up
+to 3 times, then it forces). Where a load costs 60 s, this turns a 40-minute
+batch into a 10-minute one.
+
+**Cooperative cancellation.** Long jobs report progress per phase and stop
+*between* phases — no CUDA kernels killed mid-flight.
+
+**Measures instead of guessing.** `vramd calibrate` runs a real job, samples
+VRAM per process at 20 Hz, and writes the measured footprint. Over ten real
+models on an RTX 4050, hand-written values were off by between −3154 and
++22448 MiB.
+
+## Integrating a model
+
+Three methods:
+
+```python
+from vramd.worker import WorkerAdapter, run_worker_loop
+
+
+class Adapter(WorkerAdapter):
+    name = "my-model"
+
+    def load(self, **kw):
+        import torch, mylib
+
+        return mylib.load(device=kw.get("device", "cuda"))
+
+    def generate(self, model, request):
+        if self.should_abort(request):
+            return self.cancelled_response()
+        self.report_progress(request, 0.0, "generating")
+        return {"status": "ok", "output": model(request["prompt"])}
+
+    def unload(self, model):
+        del model
+
+
+if __name__ == "__main__":
+    run_worker_loop(Adapter, backend_name="my-model")
+```
+
+And register it — without touching vramd's code:
+
+```yaml
+# ~/.config/vramd/backends.d/my-model.yaml
+version: 2
+backends:
+  - name: my-model
+    adapter: my_package.adapter
+    vram_mib: 4200
+    priority: 20
+    runtime:
+      command: ["/opt/my-model/venv/bin/python", "-m", "my_package.worker"]
+      env: { HF_HOME: ~/hf-cache }
+    load_keys: [device, compute_type]
+    shape_keys: [device]
+```
+
+Full runnable example: [`examples/echo-backend/`](examples/echo-backend/).
+
+## Calibration
+
+The friction of any such system is the "what numbers do I put in the
+descriptor?" question. vramd's answer: none — you measure.
+
+```bash
+vramd calibrate my-model --repeats 3 --out ~/.config/vramd/backends.d/measured.yaml
+```
+
+It runs the job, splits **CUDA context / weights / activation** at phase
+boundaries, and writes the descriptor. What it catches, that an estimate
+can't:
+
+| Signal | Why it matters |
+|---|---|
+| peak at **load** above the inference peak | loading fp16 and quantizing afterwards OOMs before generating |
+| activation ≫ weights | the model loads another model inside `generate` |
+| nothing resident after load | lazy loading: there's nothing to evict |
+| `unload` that doesn't return VRAM | evicting this backend frees nothing — the eviction plan would be fiction |
+| leak on repetition | the resident footprint grows with every job |
+| warmup on 1st inference | calibrating with `--repeats 1` inflates the number |
+
+Each measurement keeps the raw samples: `vramd recalibrate report.json`
+recomputes the numbers when the analysis improves, without re-occupying the
+GPU.
+
+**Calibration works out of the box for backends that need inputs.** A
+descriptor can declare the generation request and the load kwargs that
+calibration should use by default, so `vramd calibrate <backend>` works
+without flags even for backends that require inputs (`mesh_path`/`output`) or
+specific formats:
+
+```yaml
+backends:
+  - name: text3d
+    calibrate_request: { mesh_path: test-mesh.glb, output: /tmp/out.glb }
+    calibrate_load_kwargs: { compute_type: fp16 }
+```
+
+Short names that match a file bundled with the package (`test-mesh.glb`,
+`test-image.png`) are resolved to the packaged path — no test model needed.
+Load-kwargs precedence: hw-auto < descriptor < explicit.
+
+## Commands
+
+```
+start stop status queue wait cancel flush backends preload evict reap
+respawn zero stats debug bench doctor calibrate recalibrate
+```
+
+- `vramd status` / `queue` — who has the GPU and what's waiting
+- `vramd zero` — frees all idle VRAM without stopping the supervisor
+- `vramd respawn <backend>` — restarts a single worker (new code) without stopping the queue
+- `vramd doctor` — environment diagnostics
+
+**You never need `kill`.** Killing GPU processes works against the queue and
+kills the wrong workload.
+
+## Configuration
+
+```
+data/backends.yaml (example)  →  $VRAMD_BACKENDS_FILE  →  ~/.config/vramd/backends.d/*.yaml
+```
+
+Per-key overlay: a file with `{name: x, vram_mib: 5632}` fixes only that field
+and inherits the rest. That's how a calibrated descriptor takes effect without
+editing the package.
+
+Variables: `VRAMD_BACKENDS_FILE`, `VRAMD_BACKENDS_DIR`, `VRAMD_TOOLS_ROOT`,
+`VRAMD_MAX_INFLIGHT`, `VRAMD_MAX_QUEUE_DEPTH`, `VRAMD_VRAM_SAFETY_MIB`,
+`VRAMD_PRIORITY`.
+
+## Known limitations
+
+Worth knowing before adopting:
+
+- **`MAX_INFLIGHT=1` by default** — one generation at a time. The right choice
+  for 6 GB, and it underuses an A100. There is support for >1 with VRAM
+  checks, but no real *packing* yet.
+- **Multi-GPU without central placement.** `gpu_ids` is passed to the worker;
+  the supervisor neither decides placement nor accounts per device.
+- **POSIX.** Pipe reads use `select`/`O_NONBLOCK`. Windows needs a different
+  IO layer.
+- **No authentication.** Unix socket with user permissions. Local, not shared.
+- **Calibration isn't magic.** It measures what your pipeline does. A model
+  that loads everything in fp16 at once doesn't start fitting because you
+  measured it — you just learn that it doesn't fit, in 0.2 s instead of
+  mid-job.
+
+## Origin
+
+Extracted from the [AiGameKit](https://github.com/maikramer), where it was
+born to have ten generative models (text→image, →3D, →audio, →motion) share a
+6 GB RTX 4050 without manual intervention. The numbers in this README are
+measurements from that card.
+
+## Contributing
+
+[`CONTRIBUTING.md`](CONTRIBUTING.md) — getting started, style, and what this
+project values. The suite runs in ~27 s with no GPU.
+
+## License
+
+MIT — see [LICENSE](LICENSE).

vramd-0.2.4/README.md

@@ -0,0 +1,212 @@
+# vramd
+
+[![CI](https://github.com/maikramer/vramd/actions/workflows/ci.yml/badge.svg)](https://github.com/maikramer/vramd/actions/workflows/ci.yml)
+[![PyPI](https://img.shields.io/pypi/v/vramd.svg)](https://pypi.org/project/vramd/)
+[![Python](https://img.shields.io/pypi/pyversions/vramd.svg)](https://pypi.org/project/vramd/)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
+
+**VRAM admission control for generative inference on consumer GPUs.**
+
+One process holds the GPU and decides who gets in. It admits by the **real
+peak** — weights + activation + margin, not just weights — queues with
+priority and affinity, evicts by weight+LRU, and runs each model in its own
+process and venv.
+
+Built for inference that lasts **seconds to minutes** on a card that can't fit
+everything. It's not an LLM server: it doesn't optimize token throughput, it
+optimizes *fitting*.
+
+```bash
+pip install vramd          # 9 MB — the supervisor doesn't import torch
+vramd start &
+vramd submit my-model --prompt "…" --wait
+```
+
+## The problem
+
+You have 6 GB of VRAM and five models that, together, ask for 40. Each runs
+fine on its own. Together, the second job starts in the middle of the first
+and both die with `CUDA out of memory` — after the weights were already
+loaded.
+
+The usual solutions assume something that isn't true here: that the model fits
+(vLLM, TGI), that the environment is homogeneous (Ray Serve, Triton), or that
+the unit of work is a token rather than a two-minute job.
+
+## What vramd does differently
+
+**Admits by the peak, not the weights.** The question isn't "does the model
+fit" — it's "does the inference peak fit". It's the difference between
+refusing in 0.2 s and dying at 80% of the job with the weights already
+loaded.
+
+**Each model in its own venv.** A backend is a process with its own
+interpreter. Models with incompatible dependencies — torch 2.x vs 2.y,
+different CUDA wheels — coexist without seeing each other.
+
+**Affinity in the queue.** If the head of the queue needs a cold model and a
+job further back needs one already in VRAM, the scheduler skips the head (up
+to 3 times, then it forces). Where a load costs 60 s, this turns a 40-minute
+batch into a 10-minute one.
+
+**Cooperative cancellation.** Long jobs report progress per phase and stop
+*between* phases — no CUDA kernels killed mid-flight.
+
+**Measures instead of guessing.** `vramd calibrate` runs a real job, samples
+VRAM per process at 20 Hz, and writes the measured footprint. Over ten real
+models on an RTX 4050, hand-written values were off by between −3154 and
++22448 MiB.
+
+## Integrating a model
+
+Three methods:
+
+```python
+from vramd.worker import WorkerAdapter, run_worker_loop
+
+
+class Adapter(WorkerAdapter):
+    name = "my-model"
+
+    def load(self, **kw):
+        import torch, mylib
+
+        return mylib.load(device=kw.get("device", "cuda"))
+
+    def generate(self, model, request):
+        if self.should_abort(request):
+            return self.cancelled_response()
+        self.report_progress(request, 0.0, "generating")
+        return {"status": "ok", "output": model(request["prompt"])}
+
+    def unload(self, model):
+        del model
+
+
+if __name__ == "__main__":
+    run_worker_loop(Adapter, backend_name="my-model")
+```
+
+And register it — without touching vramd's code:
+
+```yaml
+# ~/.config/vramd/backends.d/my-model.yaml
+version: 2
+backends:
+  - name: my-model
+    adapter: my_package.adapter
+    vram_mib: 4200
+    priority: 20
+    runtime:
+      command: ["/opt/my-model/venv/bin/python", "-m", "my_package.worker"]
+      env: { HF_HOME: ~/hf-cache }
+    load_keys: [device, compute_type]
+    shape_keys: [device]
+```
+
+Full runnable example: [`examples/echo-backend/`](examples/echo-backend/).
+
+## Calibration
+
+The friction of any such system is the "what numbers do I put in the
+descriptor?" question. vramd's answer: none — you measure.
+
+```bash
+vramd calibrate my-model --repeats 3 --out ~/.config/vramd/backends.d/measured.yaml
+```
+
+It runs the job, splits **CUDA context / weights / activation** at phase
+boundaries, and writes the descriptor. What it catches, that an estimate
+can't:
+
+| Signal | Why it matters |
+|---|---|
+| peak at **load** above the inference peak | loading fp16 and quantizing afterwards OOMs before generating |
+| activation ≫ weights | the model loads another model inside `generate` |
+| nothing resident after load | lazy loading: there's nothing to evict |
+| `unload` that doesn't return VRAM | evicting this backend frees nothing — the eviction plan would be fiction |
+| leak on repetition | the resident footprint grows with every job |
+| warmup on 1st inference | calibrating with `--repeats 1` inflates the number |
+
+Each measurement keeps the raw samples: `vramd recalibrate report.json`
+recomputes the numbers when the analysis improves, without re-occupying the
+GPU.
+
+**Calibration works out of the box for backends that need inputs.** A
+descriptor can declare the generation request and the load kwargs that
+calibration should use by default, so `vramd calibrate <backend>` works
+without flags even for backends that require inputs (`mesh_path`/`output`) or
+specific formats:
+
+```yaml
+backends:
+  - name: text3d
+    calibrate_request: { mesh_path: test-mesh.glb, output: /tmp/out.glb }
+    calibrate_load_kwargs: { compute_type: fp16 }
+```
+
+Short names that match a file bundled with the package (`test-mesh.glb`,
+`test-image.png`) are resolved to the packaged path — no test model needed.
+Load-kwargs precedence: hw-auto < descriptor < explicit.
+
+## Commands
+
+```
+start stop status queue wait cancel flush backends preload evict reap
+respawn zero stats debug bench doctor calibrate recalibrate
+```
+
+- `vramd status` / `queue` — who has the GPU and what's waiting
+- `vramd zero` — frees all idle VRAM without stopping the supervisor
+- `vramd respawn <backend>` — restarts a single worker (new code) without stopping the queue
+- `vramd doctor` — environment diagnostics
+
+**You never need `kill`.** Killing GPU processes works against the queue and
+kills the wrong workload.
+
+## Configuration
+
+```
+data/backends.yaml (example)  →  $VRAMD_BACKENDS_FILE  →  ~/.config/vramd/backends.d/*.yaml
+```
+
+Per-key overlay: a file with `{name: x, vram_mib: 5632}` fixes only that field
+and inherits the rest. That's how a calibrated descriptor takes effect without
+editing the package.
+
+Variables: `VRAMD_BACKENDS_FILE`, `VRAMD_BACKENDS_DIR`, `VRAMD_TOOLS_ROOT`,
+`VRAMD_MAX_INFLIGHT`, `VRAMD_MAX_QUEUE_DEPTH`, `VRAMD_VRAM_SAFETY_MIB`,
+`VRAMD_PRIORITY`.
+
+## Known limitations
+
+Worth knowing before adopting:
+
+- **`MAX_INFLIGHT=1` by default** — one generation at a time. The right choice
+  for 6 GB, and it underuses an A100. There is support for >1 with VRAM
+  checks, but no real *packing* yet.
+- **Multi-GPU without central placement.** `gpu_ids` is passed to the worker;
+  the supervisor neither decides placement nor accounts per device.
+- **POSIX.** Pipe reads use `select`/`O_NONBLOCK`. Windows needs a different
+  IO layer.
+- **No authentication.** Unix socket with user permissions. Local, not shared.
+- **Calibration isn't magic.** It measures what your pipeline does. A model
+  that loads everything in fp16 at once doesn't start fitting because you
+  measured it — you just learn that it doesn't fit, in 0.2 s instead of
+  mid-job.
+
+## Origin
+
+Extracted from the [AiGameKit](https://github.com/maikramer), where it was
+born to have ten generative models (text→image, →3D, →audio, →motion) share a
+6 GB RTX 4050 without manual intervention. The numbers in this README are
+measurements from that card.
+
+## Contributing
+
+[`CONTRIBUTING.md`](CONTRIBUTING.md) — getting started, style, and what this
+project values. The suite runs in ~27 s with no GPU.
+
+## License
+
+MIT — see [LICENSE](LICENSE).

{vramd-0.2.2 → vramd-0.2.4}/pyproject.toml

@@ -4,8 +4,8 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vramd"
-version = "0.2.2"
-description = "Controlo de admissão de VRAM para inferência generativa em GPUs de consumo"
+version = "0.2.4"
+description = "VRAM admission control for generative inference on consumer GPUs"
 readme = "README.md"
 authors = [{ name = "Maikeu Locatelli" }]
 requires-python = ">=3.11"
@@ -21,8 +21,8 @@ classifiers = [
     "Topic :: System :: Distributed Computing",
 ]
 
-# Núcleo leve de propósito: o supervisor não importa torch e nem cria contexto
-# CUDA. Os modelos vivem nos workers, cada um no seu venv.
+# Deliberately light core: the supervisor doesn't import torch nor create a
+# CUDA context. Models live in the workers, each in its own venv.
 dependencies = [
     "click>=8.1.0",
     "rich>=13.0.0",
@@ -48,7 +48,7 @@ Changelog = "https://github.com/maikramer/vramd/blob/main/CHANGELOG.md"
 where = ["src"]
 
 [tool.setuptools.package-data]
-vramd = ["data/*.yaml", "py.typed"]
+vramd = ["data/*.yaml", "data/*.glb", "data/*.png", "py.typed"]
 
 [tool.pytest.ini_options]
 pythonpath = ["src", "."]

{vramd-0.2.2 → vramd-0.2.4}/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__ = "0.2.2"
+__version__ = "0.2.4"
 __all__ = ["__version__"]

{vramd-0.2.2 → vramd-0.2.4}/src/vramd/cli.py

@@ -1437,6 +1437,27 @@ def _render_comparison(rows: list[Any]) -> int:
     return under
 
 
+def _resolve_calibrate_assets(request: dict[str, Any]) -> None:
+    """Resolve inputs de teste do calibrador para o package data do vramd.
+
+    Um descriptor declara ``calibrate_request: {mesh_path: test-mesh.glb, …}``
+    com um NOME curto em vez de um path. Se o nome corresponde a um ficheiro
+    empacotado em ``vramd/data/`` (test-mesh.glb, test-image.png), é
+    substituído pelo path absoluto do pacote — o utilizador comum não precisa
+    de ter nem gerar um modelo de teste.
+    """
+    from importlib import resources
+
+    for key, value in list(request.items()):
+        if not isinstance(value, str) or not value:
+            continue
+        if Path(value).is_absolute() or "/" in value or "\\" in value:
+            continue  # path real — não mexer
+        bundled = resources.files("vramd").joinpath("data", value)
+        if bundled.is_file():
+            request[key] = str(bundled)
+
+
 @cli.command("calibrate")
 @click.argument("backend")
 @click.option("--prompt", default=None, help="Prompt do job de calibração (atalho para --request-json).")
@@ -1540,7 +1561,12 @@ def calibrate_cmd(
     load_kwargs.update(explicit_kwargs)
     if quant:
         load_kwargs.setdefault("sdnq_preset", quant)
-    quant_mode = str(load_kwargs.get("sdnq_preset") or load_kwargs.get("quant_mode") or "none")
+    # quant_preset é o alias Text2D de sdnq_preset — sem isto uma calibração
+    # com quant_preset ficava etiquetada quant_mode=none e o admit descartava
+    # a medição (quant mismatch) quando a produção pedia o preset.
+    quant_mode = str(
+        load_kwargs.get("sdnq_preset") or load_kwargs.get("quant_preset") or load_kwargs.get("quant_mode") or "none"
+    )
 
     if request_json:
         request = json.loads(Path(request_json).read_text(encoding="utf-8"))
@@ -1553,6 +1579,7 @@ def calibrate_cmd(
         if output_path:
             request["output"] = output_path
         request.setdefault("output", f"/tmp/vramd-calibrate-{backend}.bin")
+    _resolve_calibrate_assets(request)
 
     pool = SubprocessWorkerPool()
     runner = CalibrationRunner(pool)