model card: 1M context, no source-repo links, truss ships in-repo

README.md

@@ -31,7 +31,7 @@ token space.
 | Vision tower | MoonViT-3d from Kimi-K2.6, 27 layers, 1152-dim — **frozen** |
 | Projector | PatchMerger MLP (`pre_norm → linear_1 → GELU → linear_2`), 1152→4608→6144 — **trained** |
 | Image tokens | up to 4096 per image (16384 MoonViT patches, 2×2 merge) |
-| Context | 262144 |
+| Context | 1048576 (1M tokens) |
 
 ## Variants
 
@@ -66,6 +66,7 @@ python -m sglang.launch_server \
   --model-path baseten/GLM-5.2-Vision-FP8 --trust-remote-code \
   --tp-size 8 --attention-backend dsa --mm-attention-backend sdpa \
   --kv-cache-dtype fp8_e4m3 --page-size 64 \
+  --context-length 1048576 \
   --reasoning-parser glm45 --tool-call-parser glm47 \
   --port 30000
 ```
@@ -102,11 +103,11 @@ GLM-5.2 is a reasoning model: with `--reasoning-parser glm45` the chain of thoug
 
 ## Deploy on Baseten
 
-A ready-to-push Truss config lives in
-[`serve/glm5v_release/truss/`](https://github.com/basetenlabs/vision-stitching/tree/main/serve/glm5v_release/truss):
+A ready-to-push [Truss](https://truss.baseten.co) config ships in this repo under `truss/`:
 
 ```bash
-truss push --remote <your-remote>
+hf download baseten/GLM-5.2-Vision-FP8 --include 'truss/*' --local-dir ./glm5v
+cd glm5v/truss && truss push --remote <your-remote>
 ```
 
 ## How it was built
@@ -127,8 +128,9 @@ Everything except the projector is copied verbatim from public upstream releases
 - **Config** — `text_config` verbatim from the GLM release, `vision_config` describing MoonViT,
   and the projector output dim set to GLM's hidden size (6144).
 
-The assembly is fully scripted and reproducible from public repos:
-[`serve/glm5v_release/assemble.py`](https://github.com/basetenlabs/vision-stitching/blob/main/serve/glm5v_release/assemble.py).
+Nothing here needs assembling — this repo is the finished checkpoint. Every weight in it
+comes from `zai-org/GLM-5.2-FP8` (or `nvidia/GLM-5.2-NVFP4`), `moonshotai/Kimi-K2.6`, and
+`baseten/glm-5-2-projector`, so the whole thing is auditable against those public sources.
 
 ## Known issues
 
@@ -139,13 +141,13 @@ The assembly is fully scripted and reproducible from public repos:
 - **NVFP4 needs two SGLang flags**: `--disable-shared-experts-fusion` (the checkpoint quantizes
   MoE experts only, so fusion hits a 3072/6144 size mismatch) and `--disable-flashinfer-autotune`
   (the autotuner's cross-rank coordination deadlocks at MoE bring-up on B200).
-- **NVFP4 needs recent transformers.** NVIDIA's export carries `layer_types` in its text
-  config, which `transformers` below ~5.12 rejects at `AutoConfig` validation. The FP8 build
-  (from Z.ai's own export) has no such field and parses everywhere.
 - **Vision quality is projector-limited.** Only the 49.5M projector was trained; the tower and
   the language model never saw joint training. Expect solid scene description and OCR-ish
   reading, and weaker performance on tasks needing tight visual grounding.
 - First SGLang start pays a ~7-8 minute DeepGEMM JIT warmup. It is cached afterwards.
+- The full 1M context fits on 8×B200 at `--mem-fraction-static 0.85` (measured KV budget
+  ~1.16M tokens), but that budget is shared across in-flight requests — one 1M-token
+  request consumes essentially all of it. Lower `--context-length` if you want concurrency.
 
 ## License