jarrelscy/GLM-5.2-NVFP4-AQLM-hybrid-500k

Retag: vision path gated live end-to-end 2026-07-27 (full context KV, text+image correct); document vision-graft loader fix

Add GLM-5.2-Vision graft: Glm5vForConditionalGeneration config + vision head (MoonViT tower + PatchMerger projector, byte-identical BF16 from baseten) + Kimi processor + vision README. Text shards untouched.

README: note text-only modality + checkpoint parity with 1M flagship

Re-tier deep layers to 1M-flagship parity + refresh AQLM codebooks/scales

docs: point serving fork at shipped branch (GLM cudagraphs-v2 / inkling feature/inkling) + note pinned flashinfer + cutlass-dsl republish guard

Serving: 2026-07-13 promoted stack defaults + measured numbers (77/71/54 short, 49/45/36 @123K)

README: add prefill throughput (~1.2-1.6K tok/s, prefix-cached)

README: fix docker port ->8001:8001; note DCP is a stock vLLM feature

README: fork-based serving (no patching), preferred tp4-1m-mtp @ MAXLEN=500000, drop stale run_rtx6000/patched-vllm + 'speeds unverified' caveat (decode is config-bound), link SERVING.md

Add SERVING.md: current vLLM SM120 launch recipes (tp4-1m-mtp preferred)

Docker section: presets/throughput validated on 1M variant only

Add SM120 Docker deployment section

Add files using upload-large-folder tool

Add files using upload-large-folder tool

Add files using upload-large-folder tool

Add files using upload-large-folder tool

initial commit