Publish Bev v0.1.2 memory measurements and portable source builds

README.md

@@ -18,12 +18,25 @@ tags:
 
 # Bev
 
-**A ternary decision engine built around Jevfire-style one-token scoring.**
+**7.21 GB model file · 7.28 GiB measured CPU RAM · 8.30 GiB measured VRAM.**
 
-[Code & documentation](https://github.com/Reza2kn/Bev) · [Release v0.1.1](https://github.com/Reza2kn/Bev/releases/tag/v0.1.1) · [Benchmarks](https://github.com/Reza2kn/Bev/blob/v0.1.1/docs/BENCHMARKS.md)
+A ternary decision engine built around Jevfire-style one-token scoring.
+
+[Code & documentation](https://github.com/Reza2kn/Bev) · [Release v0.1.2](https://github.com/Reza2kn/Bev/releases/tag/v0.1.2) · [Benchmarks](https://github.com/Reza2kn/Bev/blob/v0.1.2/docs/BENCHMARKS.md)
 
 **The GGUF in this repository is a byte-identical redistribution of [Prism ML's Ternary-Bonsai-2-27B](https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf). Bev did not train or quantize these weights.** Prism supplies the ternary model, derived from Qwen3.8-27B. Bev adds a selected-token scoring extension, a local typed-decision API, portable setup, and measured evaluation. This is a model-and-software bundle, not a new fine-tune.
 
+## Memory and platforms
+
+| Configuration | Observed memory | Status |
+|---|---:|---|
+| Linux CPU, 1 × 4,096-token slot | **7.28 GiB peak resident system RAM** for one short selected-score inference | Tested on Stallion |
+| Linux CUDA, 2 × 16,384-token slots | **8.30 GiB GPU memory** in a serving-process snapshot | Full Persian benchmark validated on Stallion |
+| macOS Apple Silicon Metal, 1 × 4,096-token slot | **8.18 GiB sampled process RSS** on an Apple M2 with 24 GiB unified memory | **14/14 API smoke checks passed** |
+| Windows x64 CPU | System RAM required; not independently measured | Portable source-build path provided |
+
+The Mac figure is sampled process RSS, not total unified-memory pressure or a guaranteed peak. The CPU number is Linux `VmHWM` of 7,630,416 KiB, including memory-mapped model pages; loaded idle was about 7.12 GiB. The GPU number is 8,504 MiB from a separate run and is **not** a peak or a host-RAM figure. The file itself is 6.71 GiB on disk. **CPU-only recommendation: start with 16 GB system RAM. An 8 GB machine is unverified and likely too tight.** Allow headroom for the OS, API, longer contexts, and parallelism. Apple Silicon uses unified memory; RAM and Metal allocations cannot be added as separate device pools. See [installation and memory details](https://github.com/Reza2kn/Bev/blob/main/docs/INSTALL.md).
+
 ## What Bev does
 
 Provide context and finite choices. Bev evaluates one next-token distribution per field, scores every candidate, and assembles structured JSON in Python. It supports:
@@ -51,16 +64,16 @@ The runtime supports 2–255 candidates per field. The validated serving configu
 | Weights license | Apache-2.0; original LICENSE and NOTICE.txt included |
 | Code license | MIT; complete attribution in the source bundle |
 
-The `model-manifest.json` records model/runtime pins and checksums. `bev-v0.1.1-source.tar.gz` contains the complete portable source, examples, tests, runtime patch, and documentation. The Python wheel packages the API only; the source installer is needed to set up the native backend. `SHA256SUMS` covers downloadable release artifacts.
+The `model-manifest.json` records model/runtime pins and checksums. `bev-v0.1.2-source.tar.gz` contains the complete portable source, examples, tests, runtime patch, and documentation. The Python wheel packages the API only; the source installer is needed to set up the native backend. `SHA256SUMS` covers downloadable release artifacts.
 
 ## Run it
 
-The supported setup for this release is **Linux x86_64 with an NVIDIA GPU and the pinned Prism CUDA 12.8 runtime**. It was validated on an RTX 5080 Laptop GPU with 16,303 MiB total memory. The serving process used about 8,504 MiB in one observation; this is not a peak-memory measurement or a hardware minimum guarantee.
+The fully benchmarked setup is **Linux x86_64 with an NVIDIA GPU and the pinned Prism CUDA 12.8 runtime**. Portable patched-source builds are supplied for macOS Metal and Windows/Linux CPU. Use each platform’s validation status above; the same Persian accuracy numbers are not yet independently reproduced on macOS or Windows.
 
-Use the [installation guide](https://github.com/Reza2kn/Bev/blob/v0.1.1/docs/INSTALL.md) for prerequisites, then:
+Use the [installation guide](https://github.com/Reza2kn/Bev/blob/v0.1.2/docs/INSTALL.md) for prerequisites, then:
 
 ```sh
-git clone --branch v0.1.1 https://github.com/Reza2kn/Bev.git
+git clone --branch v0.1.2 https://github.com/Reza2kn/Bev.git
 cd Bev
 bash scripts/install.sh
 bash scripts/start-services.sh
@@ -72,7 +85,7 @@ curl --fail-with-body http://127.0.0.1:18781/v1/decisions \
 
 The support example returns `{"route":"billing"}` in `parsed_json`, alongside complete candidate scores. Interactive API documentation is served at `http://127.0.0.1:18781/docs`. The API binds to loopback by default.
 
-The installer verifies and downloads the original pinned Prism file. To use the identical copy from this repository instead, download it into the same model directory before installation:
+For macOS or Windows/Linux CPU, follow the [portable installation instructions](https://github.com/Reza2kn/Bev/blob/main/docs/INSTALL.md#macos-windows-and-linux-cpu). The installer verifies and downloads the original pinned Prism file. To use the identical copy from this repository instead, download it into the same model directory before installation:
 
 ```sh
 export BEV_ROOT="${BEV_ROOT:-${XDG_DATA_HOME:-$HOME/.local/share}/bev}"
@@ -83,7 +96,7 @@ This requires the Hugging Face CLI (`pip install huggingface_hub`). A generic GG
 
 ## Persian evaluation
 
-On September 23, 2026, Bev v0.1.1 ran the complete [Jev Persian Benchmark](https://github.com/ArmanJR/Jev-Persian-Benchmark) at commit `ac218d96630da9d9cc08fd897868c4d3c7048b0d`, using the original dataset, question order, batches and scorer. All **624/624 answers** were valid across **106/106 completed requests**.
+On September 23, 2026, Bev v0.1.1 (same model and scoring code as v0.1.2) ran the complete [Jev Persian Benchmark](https://github.com/ArmanJR/Jev-Persian-Benchmark) at commit `ac218d96630da9d9cc08fd897868c4d3c7048b0d`, using the original dataset, question order, batches and scorer. All **624/624 answers** were valid across **106/106 completed requests**.
 
 | Main metric | Bev | Published Jev 1.13.0 reference |
 |---|---:|---:|
@@ -98,13 +111,13 @@ Jev numbers are the benchmark author's published reference, not an independent J
 
 The measured median was **2.136 seconds per request** and total request time **224.03 seconds**. Main/repeat batches each have six questions. The hosted Jev reference and this laptop GPU have different hardware and serving conditions. No matched full-precision or ternary speedup comparison was performed.
 
-Aggregate results and provenance are included under `evaluations/`. Raw benchmark questions, gold labels, original scorer code, request journals and private host details are excluded. [Reproduction instructions](https://github.com/Reza2kn/Bev/blob/v0.1.1/docs/REPRODUCE.md) use the separately obtained upstream benchmark.
+Aggregate results and provenance are included under `evaluations/`. Raw benchmark questions, gold labels, original scorer code, request journals and private host details are excluded. [Reproduction instructions](https://github.com/Reza2kn/Bev/blob/v0.1.2/docs/REPRODUCE.md) use the separately obtained upstream benchmark.
 
 ## Limits and intended use
 
 Bev is intended for finite-label routing, classification and rubric evaluation where the application can define the allowed outputs. Fields are independent. Relative candidate probabilities are not calibrated confidence in correctness; confident mistakes occurred in evaluation.
 
-The Persian benchmark is synthetic and correlated, without independent human annotation. An earlier small general diagnostic scored **7/12 MMLU** and **2/10 SimpleBench**, alongside stronger results on other small subsets. Its loaded-source attestation was incomplete; the [full report](https://github.com/Reza2kn/Bev/blob/v0.1.1/docs/BENCHMARKS.md) retains this limitation. Neither run establishes broad reliability, Jev parity, or a Decision Index rank.
+The Persian benchmark is synthetic and correlated, without independent human annotation. An earlier small general diagnostic scored **7/12 MMLU** and **2/10 SimpleBench**, alongside stronger results on other small subsets. Its loaded-source attestation was incomplete; the [full report](https://github.com/Reza2kn/Bev/blob/v0.1.2/docs/BENCHMARKS.md) retains this limitation. Neither run establishes broad reliability, Jev parity, or a Decision Index rank.
 
 One-token scoring can miss problems requiring multi-step reasoning, and the model inherits limitations and biases from its upstream models. A constrained output format does not guarantee a correct decision. No new calibration or independent production-domain validation is supplied by this release.
 

SHA256SUMS

@@ -1,13 +1,18 @@
 69849221bfb90053de2134ef5e6d540287b4b98062326492f1f96f5da685524b  LICENSE
 9d75bb94ffefed09e8060f4cf54313921b91fa11777eb3e3977cd6e42672b478  LICENSE-CODE.txt
 de0e0c48fb6f691a31e74f338e3ccf93f9ecdfe2866ab769c4bf8b79a7636a30  NOTICE.txt
-d79f0e5b7cbba306b57cb049e99f585b3da43f71d9570e5f98aa38903015746e  README.md
+1764a54d25c635001095dacefdaae157907574bf50fe9e9e6bd61760b2465953  README.md
 3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1  Ternary-Bonsai-2-27B-PQ2_0.gguf
 817c2308bebadd29eb4ff465ff3e540c0926f3652194db820472ef0c038250bc  bev-v0.1.1-source.tar.gz
+0211ae127f11200e87fa0b0f79a4380f820093ab953a7b283742ba08fc0dc082  bev-v0.1.2-source.tar.gz
 39bbdc952792a6047ea853ce6d2494a475bddbe56291c5bf218c0a680f8d4361  bev_decisions-0.1.1-py3-none-any.whl
 77e0dc9697310a960fcc080ea48189b9bc050dd73077497225235f11a8ab85e4  bev_decisions-0.1.1.tar.gz
+999b6275fa0c6fba42c94d421c81e16d4af22126dd2f6a0219f5cd3e7819adfd  bev_decisions-0.1.2-py3-none-any.whl
+2ad8e3dba5525cf6e37e494098d7280d02d85b081275d4232fa22b4de2589b06  bev_decisions-0.1.2.tar.gz
 f23b48a5244b8a6a2ad9f9b8fe6df6393ec0083e5c87c2e9fa1ba754d23ba35d  evaluations/general-diagnostic-v1/provenance.json
 3cd50061c268d3ed3b6cd14e6de42a1802b578349367c2fe5c9b7120727e2305  evaluations/general-diagnostic-v1/summary.json
+f2677b5ebfe59a840409d4a3c0bcd3e475177d9d09caff3d5266eddf57930d88  evaluations/memory-cpu-stallion.json
+3f9b479e095553046f2351a1f3e1070bd7106b5e412f802956c614d1f756c78f  evaluations/memory-macos-metal.json
 42b25662be940fe9048e5324ca05fd80267efcf43ec648c9c5365f7b09ac5657  evaluations/packaging-validation.json
 c55f14408e5a2fb6e7703549aaa92a832c54b63ea4208bbd378f2032bf8d47b9  evaluations/persian-v1/provenance.json
 b2403e483fef2905e5bba0333041ff48fdf2d385979249dc8693ea3da9d6a0cc  evaluations/persian-v1/runner.py
@@ -18,5 +23,5 @@ d658541cffbaa2c53b45920330e3ae5a7e461c580c80b12f8b026ac30672c0ac  evaluations/wh
 69849221bfb90053de2134ef5e6d540287b4b98062326492f1f96f5da685524b  licenses/MODEL-APACHE-2.0.txt
 de0e0c48fb6f691a31e74f338e3ccf93f9ecdfe2866ab769c4bf8b79a7636a30  licenses/MODEL-NOTICE.txt
 94f29bbed6a22c35b992c5c6ebf0e7c92f13b836b90f36f461c9cf2f0f1d010d  licenses/PRISM-LLAMA-MIT.txt
-a3fdeeed74218bfa11deb50de2d1791f29615b9f10d1fdcad90b67b6ab7ccaec  model-manifest.json
-9ee4b238adced4bb401b4c45666f90509bdb78a30452c8c41c9d11b64b08edea  release-manifest.json
+e9b2944faf5575b515d979fe5c426cf59480f4768cfc29bc8437bb3befb35b2c  model-manifest.json
+fd942325cd5d879c3062b3d38c18227115f392c3390e923cd1f72d54701dadfe  release-manifest.json

bev-v0.1.2-source.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0211ae127f11200e87fa0b0f79a4380f820093ab953a7b283742ba08fc0dc082
+size 96748

bev_decisions-0.1.2.tar.gz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2ad8e3dba5525cf6e37e494098d7280d02d85b081275d4232fa22b4de2589b06
+size 102356

evaluations/memory-cpu-stallion.json

@@ -0,0 +1,19 @@
+{
+  "date_utc": "2026-09-23",
+  "host": "Stallion, Linux x86_64",
+  "runtime_commit": "9a9394a895b96003ca842a6041cb28ac49a108f7",
+  "model_sha256": "3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1",
+  "backend": "CPU only, n_gpu_layers=0",
+  "context_total_tokens": 4096,
+  "parallel_slots": 1,
+  "kv_type": "q8_0",
+  "request": "One short /completion request with selected_token_ids=[1,2] and n_predict=1; response returned HTTP 200 with one token.",
+  "loaded_idle_vm_rss_kib": 7468732,
+  "post_inference_vm_hwm_kib": 7630416,
+  "post_inference_vm_rss_kib": 7630416,
+  "post_inference_pss_kib": 7626623,
+  "post_inference_pss_file_kib": 7037583,
+  "post_inference_pss_anon_kib": 589040,
+  "measurement_source": "/proc/<owned Bev server PID>/status and smaps_rollup, immediately after the request",
+  "interpretation": "RSS includes file-backed resident model pages; this is an observed run, not a general minimum or controlled peak across arbitrary requests."
+}

evaluations/memory-macos-metal.json

@@ -0,0 +1,15 @@
+{
+  "date_utc": "2026-09-23",
+  "host": "Apple M2 Mac with 24 GiB unified physical memory",
+  "runtime_commit": "9a9394a895b96003ca842a6041cb28ac49a108f7",
+  "model_sha256": "3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1",
+  "backend": "Metal; n_gpu_layers=99",
+  "context_total_tokens": 4096,
+  "parallel_slots": 1,
+  "kv_type": "q8_0",
+  "functional_smoke_passed": 14,
+  "functional_smoke_total": 14,
+  "observed_max_rss_kib": 8580912,
+  "measurement_source": "macOS ps rss sampled once per second during API smoke, including long prompt and repeated decisions",
+  "interpretation": "Observed process RSS only, not total unified-memory pressure or a guaranteed peak; does not establish 8 GB device viability."
+}

model-manifest.json

@@ -1,6 +1,6 @@
 {
   "name": "Bev",
-  "version": "0.1.1",
+  "version": "0.1.2",
   "model_id": "prism-ml/Ternary-Bonsai-2-27B-gguf",
   "model_revision": "6ed5e12bf84b7a63069882c91dd9e9218647d17b",
   "model_file": "Ternary-Bonsai-2-27B-PQ2_0.gguf",
@@ -25,5 +25,10 @@
   "code_license": "MIT",
   "github": "https://github.com/Reza2kn/Bev",
   "huggingface": "https://huggingface.co/Reza2kn/Bev",
-  "weight_relationship": "Byte-identical redistribution of the pinned Prism GGUF; no new training or quantization."
+  "weight_relationship": "Byte-identical redistribution of the pinned Prism GGUF; no new training or quantization.",
+  "system_ram_cpu_kib_peak_observed": 7630416,
+  "system_ram_cpu_configuration": "Linux CPU; one 4096-token slot; one short selected-score request; VmHWM; includes file-backed resident pages",
+  "model_vram_mib_observation_type": "Linux CUDA serving-process snapshot; not a measured peak",
+  "macos_metal_sampled_max_rss_kib": 8580912,
+  "macos_metal_configuration": "Apple M2 with 24 GiB unified RAM; 4096-token single slot; 14/14 API smoke checks"
 }

release-manifest.json

@@ -1,61 +1,20 @@
 {
-  "name": "Bev",
-  "version": "0.1.1",
-  "release_date": "2026-09-23",
-  "git_commit": "c5c12471a6eab4fd31da30b3d442f1f6dc2bd503",
-  "github": "https://github.com/Reza2kn/Bev",
-  "huggingface": "https://huggingface.co/Reza2kn/Bev",
-  "source_files": 52,
-  "artifacts": {
-    "bev_decisions-0.1.1-py3-none-any.whl": {
-      "bytes": 22141,
-      "sha256": "39bbdc952792a6047ea853ce6d2494a475bddbe56291c5bf218c0a680f8d4361"
+  "version": "0.1.2",
+  "git_commit": "be5d575baec7004a0b3fb6eee808c5ddbf922c43",
+  "model_sha256": "3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1",
+  "model_bytes": 7206168928,
+  "files": {
+    "bev-v0.1.2-source.tar.gz": {
+      "size": 96748,
+      "sha256": "0211ae127f11200e87fa0b0f79a4380f820093ab953a7b283742ba08fc0dc082"
     },
-    "bev-v0.1.1-source.tar.gz": {
-      "bytes": 88732,
-      "sha256": "817c2308bebadd29eb4ff465ff3e540c0926f3652194db820472ef0c038250bc"
+    "bev_decisions-0.1.2-py3-none-any.whl": {
+      "size": 22664,
+      "sha256": "999b6275fa0c6fba42c94d421c81e16d4af22126dd2f6a0219f5cd3e7819adfd"
     },
-    "bev_decisions-0.1.1.tar.gz": {
-      "bytes": 91774,
-      "sha256": "77e0dc9697310a960fcc080ea48189b9bc050dd73077497225235f11a8ab85e4"
+    "bev_decisions-0.1.2.tar.gz": {
+      "size": 102356,
+      "sha256": "2ad8e3dba5525cf6e37e494098d7280d02d85b081275d4232fa22b4de2589b06"
     }
-  },
-  "checks": {
-    "unit_and_package_tests": 53,
-    "curated_source_checks_passed": true,
-    "native_installer_fresh_build_passed": true,
-    "documented_examples_passed": true,
-    "clean_wheel_api_passed": true,
-    "inference_source_matches_persian_baseline": true
-  },
-  "build_compute": "Linux x86_64; all builds, tests and model inference performed on the original NVIDIA GPU host",
-  "model": {
-    "name": "Bev",
-    "version": "0.1.1",
-    "model_id": "prism-ml/Ternary-Bonsai-2-27B-gguf",
-    "model_revision": "6ed5e12bf84b7a63069882c91dd9e9218647d17b",
-    "model_file": "Ternary-Bonsai-2-27B-PQ2_0.gguf",
-    "model_bytes": 7206168928,
-    "model_sha256": "3907dc1658db1f78a9826bf8d5bcb8dc65db0d466388937af57f2294fae62ec1",
-    "packing": "PQ2_0, ternary weights with group scaling in two-bit slots",
-    "weights_modified": false,
-    "training_performed": false,
-    "jevfire_revision": "5df83b558bf635e006d31f8f2fc2798d0e3ff051",
-    "runtime_repository": "https://github.com/PrismML-Eng/llama.cpp",
-    "runtime_revision": "9a9394a895b96003ca842a6041cb28ac49a108f7",
-    "runtime_release": "prism-b10709-9a9394a",
-    "gpu": "NVIDIA GeForce RTX 5080 Laptop GPU",
-    "vram_mib": 16303,
-    "context_tokens_per_slot": 16384,
-    "parallel_slots": 2,
-    "runtime_patch_sha256": "7dadeaf5fccc2cb19c7c78e76173368168f9c904d7c88a6afe844722bba91cc2",
-    "runtime_archive_sha256": "8aec67eb023b251712c7e6490f367b5671bf587eced1436a9b85f4a90c3b7d3d",
-    "cuda_library_sha256": "a2f67b3e1a3fb476aba8b6fc1d2d2d7b25add3c4a169d77a679288492acc3762",
-    "model_vram_mib": 8504,
-    "weights_license": "Apache-2.0",
-    "code_license": "MIT",
-    "github": "https://github.com/Reza2kn/Bev",
-    "huggingface": "https://huggingface.co/Reza2kn/Bev",
-    "weight_relationship": "Byte-identical redistribution of the pinned Prism GGUF; no new training or quantization."
   }
 }