Remove obsolete v3 root Compose

Publish validated v20 profiles, calibrated KV scales, and release card

v3.6: damage-optimal K64 + BF16-stored ne-tier (GPQA 88.89, KLD 0.126 vs BF16-ref); 184x<=2GB shards; compose = measured 422k-pool config

Add files using upload-large-folder tool

Add files using upload-large-folder tool

Add files using upload-large-folder tool

Serving image v3: baked AR-crossover tuning (CARPATCH) + ONESHOTMAX=65536 -> public decode 95-101 t/s; dogfooded on the reference box (pool 428,544, gates green)

Daily-driver envelope: 262k context, 428,544-token KV pool (2048-chunk squeeze @ util 0.968)

Hotfix: safetensors load (BF16-fixed checkpoint OOMs fastsafetensors staging)

Dense-tier BF16 fix: kv_b_proj + shared_experts now stored as true BF16 (zai-original bytes, layout identical to lukealonso/GLM-5.2-NVFP4); their fp8 weight_scale tensors removed. Measured KLD-neutral at 2k context (0.211 vs 0.209, 5-run BF16-reference protocol); motivated by attention-path correctness at long context, convention alignment, and observed codegen quality. Serving note: fastsafetensors staging may load-OOM on the larger checkpoint at very high gpu-memory-utilization; plain safetensors loads fine.

240k context, rev-3 KLD 0.177

maxlen 240k (fits single-request)

serving: 300k context default, 301k pool

max-model-len 300k for usable long context

Align compose to exact GPQA-run config (maxseqs 8 / capsize 64)

decode 60-90 t/s

KLD section: vs-FP8 table + vs-BF16 frame caveat

Fix strikethrough rendering (tildes -> approx symbol)

Add daily-driver docker-compose (v2 image, 307k pool)

Update serving: v2 image, 307k KV pool, mtp5/dcp4/fastsafetensors config

Add model card: GPQA-Diamond 88.38 (parity with full precision), NF3-hybrid + mxfp8 tier

Add files using upload-large-folder tool

Add files using upload-large-folder tool

initial commit