Fix Space config: restore Docker SDK Dockerfile + frontmatter README

The source-tree copy had clobbered both files (generic remote Dockerfile, CISM README without sdk frontmatter). This restores the Space entrypoint (fleet then serve /results on 7860) and the sdk: docker frontmatter.

Dockerfile

@@ -1,20 +1,12 @@
-# CISM remote-autobench image: builds the native extension and runs the
-# full AutoBench matrix (all precisions x threads x spec K + HF baselines)
-# over a fleet of models. Results land in /results (mount a volume).
-#
-# Build:   docker build -t cism-bench .
-# Full:    docker run --rm -v $(pwd)/results:/results cism-bench
-# Quick:   docker run --rm -v $(pwd)/results:/results cism-bench --quick
-# Custom:  docker run --rm -v $(pwd)/results:/results cism-bench \
-#            --models SupraLabs/Supra-Mini-v5-8M --out-dir /results --reps 1
-#
-# Notes:
-# - torch is installed CPU-only first (PyPI default would pull CUDA wheels);
-#   the [reference] extra is then already satisfied and HF baselines run.
-# - The full 4-model fleet takes hours (inductor warmups + 72 cells x reps=3
-#   per model). Use --quick for a smoke run (~minutes on one small model).
-# - Models download from the Hub on first run; mount an HF cache to reuse:
-#   -v ~/.cache/huggingface:/root/.cache/huggingface
+# CISM remote-autobench Space (CPU): builds the native extension, runs the
+# full AutoBench fleet (all precisions x threads x spec K + HF baselines),
+# then serves /results on 7860 so numbers stay browsable while the Space
+# lives. Fleet selection without rebuild via Space Variables:
+#   CISM_MODELS  (space-separated HF ids; default: 4-model fleet)
+#   CISM_OUT_DIR (default /results)
+# Optional Secret for off-Space backup:
+#   HF_TOKEN     (enables --upload-to below)
+#   CISM_UPLOAD_TO (dataset repo id; default spitfire4794/test1111111-results)
 
 FROM ubuntu:24.04
 
@@ -32,17 +24,16 @@ RUN python3 -m venv /opt/venv \
 
 WORKDIR /cism
 
-# Build deps first for layer caching; source last for fast rebuilds.
 COPY pyproject.toml README.md CMakeLists.txt ./
 COPY native/ native/
 COPY src/ src/
 COPY scripts/autobench_fleet.py scripts/autobench_fleet.py
+COPY entrypoint.sh ./
 
 RUN pip install torch --index-url https://download.pytorch.org/whl/cpu \
     && pip install ".[test,server]" \
     && python -c "from cism import _native as n; print('kernel_variant:', n.kernel_variant())"
 
-VOLUME ["/results"]
+EXPOSE 7860
 
-ENTRYPOINT ["python", "scripts/autobench_fleet.py"]
-CMD ["--out-dir", "/results"]
+ENTRYPOINT ["bash", "entrypoint.sh"]

README.md

@@ -1,193 +1,26 @@
-# CISM
-
-Experimental native CPU inference with a Python SDK and OpenAI-compatible text
-server. Private development snapshot (`0.1.0.dev0`), **not the v0.1 beta yet**.
-No performance advantage over llama.cpp has been established.
-
-## Implemented
-
-- Real C++20 autoregressive Llama/Qwen3 inference: RMSNorm, RoPE, GQA, SwiGLU,
-  residuals, incremental KV, native greedy/top-k/top-p sampling.
-- Runtime-dispatched AVX2 FP32, weight-only INT8 and hybrid block-32 INT4 kernels,
-  with scalar fallbacks. Activations remain FP32 in every mode.
-- Direct HF/local safetensors import, including BF16 conversion without PyTorch.
-- Python generation and bounded FIFO HTTP serving, SSE, usage, cancellation,
-  optional bearer authentication, and strict unsupported-field errors.
-- One model and one active generation; single-threaded native execution.
-
-## Build on Windows
-
-Install **Visual Studio 2022 Build Tools**, workload **Desktop development with
-C++**, including MSVC v143 x64/x86, Windows SDK, and C++ CMake tools.
-Use **Developer PowerShell for VS 2022**:
-
-```powershell
-python -m venv .venv
-.venv\Scripts\python -m pip install ".[server,test]"
-.venv\Scripts\python -m pytest
-```
-
-Python 3.11-3.13 is targeted. Build instructions above compile locally; prebuilt
-wheels are not distributed yet. No CUDA, MinGW, or full Visual Studio IDE needed.
-
-## Build on Linux
-
-Install a C++20 compiler and Python development headers, then:
-
-```bash
-python3 -m venv .venv
-.venv/bin/python -m pip install '.[server,test]'
-.venv/bin/python -m pytest
-```
-
-Linux support is intended but must be validated on a Linux host before beta.
-
-## Use
-
-```python
-from cism import Engine
-
-engine = Engine.from_pretrained("SupraLabs/Supra-Mini-v5-8M", precision="int8")
-print(engine.info)
-print(engine.generate("The future of artificial intelligence is", max_new_tokens=32).text)
-```
-
-```text
-cism inspect SupraLabs/Supra-Mini-v5-8M
-cism generate SupraLabs/Supra-Mini-v5-8M --precision int8 --max-tokens 32
-cism bench SupraLabs/Supra-Mini-v5-8M --precision fp32 --runs 5
-cism serve SupraLabs/Supra-Mini-v5-8M --precision int8
-```
-
-`--threads N` (all commands) runs decode with a persistent native worker pool;
-2 threads is typically the sweet spot for cache-sized quantized models on
-desktop Zen 3 and is exposed per command. Matrix sizes below the parallel
-threshold stay on the calling thread regardless.
-
-`cism generate/bench --spec K` enables greedy prompt-lookup speculation
-(K in 2-8, requires temperature=0): an n-gram drafter proposes continuations
-from the token history and a GEMM verify pass accepts them in one weight
-stream. Output is identical to plain greedy decode. It pays off when the model
-streams from DRAM (larger than L3) and the text repeats prompt patterns; it
-gains little on cache-resident models or free-form text.
-
-`serve` accepts several models and routes requests by the `model` field through
-one shared FIFO worker, so only one generation runs at a time:
-
-```text
-cism serve SupraLabs/Supra-Mini-v5-8M SupraLabs/Supra2-Medium-Base HuggingFaceTB/SmolLM2-135M --precision int8
-```
-
-`--revision` pins a HF revision; resolved commits are reported in engine info.
-Use `--local-files-only` offline and `--trust-remote-code` only for trusted custom
-tokenizers. Native architecture support is never inferred from permission to run
-Python. Import caches use the HF snapshot cache; native packing currently repeats
-at model load rather than persisting a compiled artifact.
-
-## HTTP
-
-Install the `server` extra. Default bind address: `127.0.0.1:8000`.
-Set `CISM_API_KEY` to require `Authorization: Bearer <key>`; unset/empty disables
-authentication. Run exactly one Uvicorn process.
-
-```python
-from openai import OpenAI
-
-client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="unused")
-result = client.completions.create(
-    model="SupraLabs/Supra-Mini-v5-8M",
-    prompt="The future of artificial intelligence is",
-    max_tokens=32,
-)
-print(result.choices[0].text)
-```
-
-Endpoints: `/v1/completions`, `/v1/chat/completions`, `/v1/models`, `/health`,
-`/ready`. Chat requires a real tokenizer chat template; base models without one
-must use completions. Streaming supports `stream_options.include_usage`.
-`--queue-size` bounds active plus waiting/unconsumed requests (default 8).
-Unsupported request fields return an error; tools, JSON constraints, logprobs,
-embeddings, and Responses are **not implemented in this snapshot**.
-
-Over-budget `max_tokens` is clamped: generation runs until the model's context
-window fills and returns `finish_reason="length"` with accurate usage. Only a
-prompt that alone exceeds the model context is rejected. This limit is the
-checkpoint's trained `max_position_embeddings`, not a server preference.
-
-Stop strings are excluded from returned text. Usage counts sampled tokens,
-including stopping/EOS tokens. Chunked generation crosses the Python boundary in
-groups, except when stop strings require checking every token. Partial UTF-8 is
-held back during streaming. Custom tokenizers that rewrite already emitted text
-are explicitly rejected. Cancellation is cooperative; a matrix operation or
-allocation cannot be interrupted mid-operation.
-
-## Web Tester
-
-The server serves a dependency-free test page at `http://127.0.0.1:8000/`.
-Pick any loaded model from `/v1/models`, choose completions or chat, set
-sampling parameters, and stream with live TTFT and tokens/s. Paste the API key
-if `CISM_API_KEY` is set; the page stores it only in your browser if you ask,
-and the server never embeds it in the HTML. All-whitespace results are labeled
-as greedy space-token loops.
-
-## Memory And Cache Control
-
-x86 CPUs have no user-controllable way to pin L1/L2/L3 cache lines. While
-generating, weights stay cache-hot because the decoder rereads them constantly.
-During idle gaps, Windows, background processes, and other workloads evict
-lines, and nothing can forbid that. CISM therefore provides two real controls:
-
-- **Page locking.** `cism serve --lock-pages` (or `Engine.lock_pages()`) pins
-  weight memory in physical RAM with `VirtualLock`/`mlock`, so the OS cannot
-  page model memory to disk until `Engine.unlock_pages()`. The return value is
-  the byte sum of successful OS calls; neighboring regions share pages, so
-  lock/unlock counts are approximate. Verified on Windows: 8,022,880 of 8,022,784
-  bytes locked for Supra-Mini-8M INT8.
-- **Keep-warm sweeps.** `cism serve --keep-warm SECONDS` (or
-  `Engine.start_keep_warm(SECONDS)` / `stop_keep_warm()`) re-reads every weight
-  byte on an interval while the engine is idle, refreshing cache residency.
-  Sweeps pause during generation and stop when told; after stopping, hardware
-  eviction proceeds naturally. `Engine.touch()` runs one sweep on demand.
-
-These reduce cold-start after idle; they do not reserve L3 or guarantee SRAM
-residency, and their effect should be measured, not assumed.
-
-### Cache Report
-
-`cism cache-report MODEL` measures whether the cache actually matters for your
-model on this machine: a warm full-weight scan, a cold scan after evicting
-caches with a scrub buffer (`--scrub-mb`, about twice the L3 size), and the
-decode effective bandwidth (weight bytes x decode tok/s). It reports the
-warm/cold ratio, the residency verdict, and the decode bottleneck verdict
-(compute-bound, DRAM-streaming, or cache-bandwidth-limited). Definitive
-per-event counters (LLC misses) require external PMU tooling such as
-WindowsPerf or AMD uProf.
-
-## Limits
-
-Only standard dense `LlamaForCausalLM` and `Qwen3ForCausalLM` with full attention,
-default split-half RoPE, SiLU MLP, and no biases are supported. Model-config limits
-are enforced; long-context extensions are not inferred. Original custom targets
-(Negative, CMA, Ember, Spark, Blaze) remain beta work, not Python fallbacks.
-
-Prefill is currently sequential, and there is no thread autotuner or persistent
-native pool. Weights, scales, KV,
-and scratch all compete for cache. CISM does not guarantee SRAM residency or
-reserve L3. The original weight-loop benchmark is not a decoder throughput test.
-
-See [the release plan](docs/ROADMAP.md) and [native details](native/README.md).
-
-## Reference Validation
-
-Install the `reference` extra to compare against PyTorch/Transformers on CPU:
-
-```text
-python scripts/validate_reference.py SupraLabs/Supra-Mini-v5-8M --precision fp32
-python scripts/validate_reference.py SupraLabs/Supra-Mini-v5-8M --precision int8
-python scripts/validate_reference.py veyra-ai/Veyra2-Blueberry-10M-Base --precision fp32
-```
-
-Quantized numerical parity uses a reference reconstructed from the exact native
-quantization definition, not the original FP32 checkpoint. This tests kernel
-correctness, not preservation of model quality. Held-out quality reports and
-matched llama.cpp benchmarks remain release gates.
+---
+title: Test1111
+emoji: 📈
+colorFrom: red
+colorTo: yellow
+sdk: docker
+pinned: false
+---
+
+# CISM remote autobench
+
+This Space builds the CISM native CPU inference engine and runs the full
+AutoBench fleet on Space CPU: all four CISM precisions
+(fp32/int8/hybrid-int4/hybrid-fp4) × thread sweep × spec-K sweep with
+median-of-3 rigor + HF eager / dynamic-int8 / torch.compile baselines.
+
+While the fleet runs, container logs stream per-cell progress. When each
+model finishes, `<model>.json` plus `SUMMARY.md` appear here (this page serves
+`/results`). Configure without rebuild via **Settings → Variables**:
+
+- `CISM_MODELS` — space-separated HF ids (default: Supra Mini-8M, Supra-50M,
+  Surjo-50m, FWKV Myosotis-1-base)
+- `CISM_OUT_DIR` — default `/results`
+
+Optional **Secret** `HF_TOKEN` (+ Variable `CISM_UPLOAD_TO`) backs results up
+to a dataset repo after the run.