brandonmusic/GLM-5.2-EXL3-TR3v4-3.5bpw-MTP78

docs: add reviewed attribution and credits

Correct r17 CUDA graph claim and mark eager launcher provisional

Make validated eager G64 runtime the download default

Fix r17 max-batched-token environment name

Make sealed r17 exact-G64-Q-only reader the download default

Document exact G64 Q-only five-run KLD pass

Document sealed r17 FP8 reader KLD comparison

Accuracy battery (LAVD/Estonia/Hotel/needle) on v39_502k, KV 502016

Add max-KV serving profile: 502,016 KV tokens @262K, MTP-3/DCP4 + prefill/decode results

Add max-KV serving profile: 502,016 KV tokens @262K, MTP-3/DCP4 + prefill/decode results

Add max-KV serving profile: 502,016 KV tokens @262K, MTP-3/DCP4 + prefill/decode results

Add max-KV serving profile: 502,016 KV tokens @262K, MTP-3/DCP4 + prefill/decode results

Model card: Gilded Gnosis r34 qualification — canonical run profile, release-gate results (credit: Festr and the local-inference-lab Discord)

Document and validate BF16-shared online K6 revision (#3)

Shared experts: BF16 + merged online K6 (split-payload decode loss isolated by Festr)

Model card: fused image published as verdictai/glm52-exl3-sparkinfer:v39-r28-r7fused-broadcast-cu132-sm120a

Point SERVING_FUSED.md at the published fused image

Point docker-compose.yaml at the published fused image

Point serve.sh at the published fused image

Model card: add fused MoE path (KV 355,328 @262K ctx, KLD 0.069527), document the required nvfp4 outer-scales file, and record the FC2-bound correctness fix

Add fused-MoE patches for custom image builds

Add fused-MoE patches for custom image builds

Add fused-MoE patches for custom image builds

Add fused-MoE patches for custom image builds

Add fused-MoE patches for custom image builds

Add fused-MoE patches for custom image builds

Add fused-MoE patches for custom image builds

Add SERVING_FUSED.md: fused MoE serving (KV 355,328 @262K ctx, KLD 0.069527)

Add docker-compose.yaml: fused MoE serving (KV 355,328 @262K ctx, KLD 0.069527)

Add serve.sh: fused MoE serving (KV 355,328 @262K ctx, KLD 0.069527)

Add nvfp4_mla_outer_scales.json: fused MoE serving (KV 355,328 @262K ctx, KLD 0.069527)

Model card: add measured prefill route-block-size tuning (+44.7% prefill, +21.8% decode) and the levers that did not help

SERVING.md: quote the measured KV capacity with its config instead of the auto-profile ceiling

Model card: KV cache capacities are now measured, with util/dtype/ctx/batch stated; qualify the 1.13M auto-profile figure

Model card: repair corrupted Run it block; link image, compose, serve script and results

Model card: surface Docker Hub image, compose, serve script and results up top

Model card: add KV cache section (supported dtypes + measured capacity) and link SERVING/RESULTS

Add RESULTS.md: serving (Docker Hub image + compose) and measured KLD/throughput results

Add SERVING.md: serving (Docker Hub image + compose) and measured KLD/throughput results

corrected r7 manifest: r7-experts-layer-077.json

corrected r7 experts: r7-experts-layer-077.safetensors

corrected r7 manifest: r7-experts-layer-076.json

corrected r7 experts: r7-experts-layer-076.safetensors

corrected r7 manifest: r7-experts-layer-075.json

corrected r7 experts: r7-experts-layer-075.safetensors

corrected r7 manifest: r7-experts-layer-074.json

corrected r7 experts: r7-experts-layer-074.safetensors

corrected r7 manifest: r7-experts-layer-073.json

corrected r7 experts: r7-experts-layer-073.safetensors

corrected r7 manifest: r7-experts-layer-072.json