README: architecture, split-part download, 0.2.4 requirement, DSA opt-in, M3 Ultra measurements

README.md

@@ -4,16 +4,18 @@ base_model: zai-org/GLM-5.2
 library_name: basert
 pipeline_tag: text-generation
 tags:
-  - basert
-  - apple-silicon
-  - quantized
+- basert
+- apple-silicon
+- quantized
 ---
 
 # GLM-5.2
 
-[BaseRT](https://github.com/basecompute/baseRT) `.base` builds of [`zai-org/GLM-5.2`](https://huggingface.co/zai-org/GLM-5.2) for fast local inference on Apple Silicon (Metal).
+[BaseRT](https://github.com/basecompute/baseRT) `.base` build of [`zai-org/GLM-5.2`](https://huggingface.co/zai-org/GLM-5.2) for local inference on Apple Silicon (Metal).
 
-A 78-layer MoE reasoning model with DSA sparse attention and an MTP layer. Q4 needs 512 GB Apple Silicon.
+A 78-layer mixture-of-experts reasoning model: 256 routed experts with 8 active per token plus one always-on shared expert (sigmoid gating), Multi-head Latent Attention (MLA, 64 heads over a single 576-wide latent KV row) with DeepSeek-style sparse attention (DSA), a 155k vocabulary and a 1M-token context window. The first three layers use a dense FFN.
+
+Needs 512 GB of unified memory: the Q4 weights alone are ~418 GB resident. Apple Silicon only — the MLA/DSA kernels have no CUDA build.
 
 ## Files
 
@@ -21,21 +23,45 @@ Q4 is 432 GB, past the Hub's per-file ceiling, so it ships as 9 parts of ≤45 G
 
 | File | Precision | Size |
 |------|-----------|------|
-| `parts/GLM-5.2-Q4.base.part-000` … `part-008` | 4-bit | 432 GB total |
+| `parts/GLM-5.2-Q4.base.part-000` … `part-008` | 4-bit (group-quantized, f16 attention norms / MLA projections) | 432 GB total |
 
-Reassemble before use:
+Reassemble before use. The parts concatenate in order into one file; stream them so the download never needs a second 432 GB of free space:
 
 ```sh
-cat parts/GLM-5.2-Q4.base.part-* > GLM-5.2-Q4.base
+for i in 0 1 2 3 4 5 6 7 8; do
+  curl -L -H "Authorization: Bearer $HF_TOKEN" \
+    "https://huggingface.co/basecompute/GLM-5.2/resolve/main/parts/GLM-5.2-Q4.base.part-00$i" >> GLM-5.2-Q4.base
+done
 shasum -a 256 GLM-5.2-Q4.base   # expect e7cf483188806908f290a6eed1d96ec6a974cb3d4751a867d5c67a317646c230
 ```
 
 ## Usage
 
+GLM 5.2 needs BaseRT 0.2.4 or newer. `basert pull` does not yet reassemble split bundles, so point the CLI at the reassembled file:
+
 ```bash
 curl -LsSf https://basecompute.co/install.sh | sh
-basert pull basecompute/GLM-5.2
-basert chat basecompute/GLM-5.2
+basert chat ./GLM-5.2-Q4.base
+basert serve ./GLM-5.2-Q4.base        # OpenAI-compatible HTTP server
+```
+
+DSA sparse attention is opt-in. Dense MLA is the default and matches llama.cpp's glm-dsa implementation bitwise; set `BASERT_DSA=1` to decode on the sparse top-k path once the context is deeper than the indexer's selection width (2048 tokens):
+
+```bash
+BASERT_DSA=1 basert serve ./GLM-5.2-Q4.base
 ```
 
+## Measured
+
+Mac Studio M3 Ultra (80-core GPU, 512 GB), BaseRT 0.2.4 (staging build of commit 36bd7e24), greedy, single stream, `basert bench` (pp = prompt tokens/s, tg = generated tokens/s):
+
+| test | dense MLA | `BASERT_DSA=1` |
+|---|---:|---:|
+| pp512 | 236 t/s | — |
+| tg128 | 22.9 t/s | — |
+| pp4096 | 178 t/s | 188 t/s |
+| tg128 after 4096 | 22.7 t/s | 22.4 t/s |
+
+Weights resident: ~418 GB. First load after a reboot streams the whole file from disk (about 3.5 minutes on the Studio's internal SSD); a warm load is under 10 seconds. At a 4k context the sparse path is a wash — its selection width is 2048 tokens, so it only starts to pay past that, and generation is weight-bandwidth-bound at this size either way.
+
 Released under the mit license, inherited from the base model.