docs: add reviewed attribution and credits

Update README.md

RELEASE_TEST_SUITE.md: pin v31 (base refresh sic3828fd; wheel superset-verified)

README.md: pin v31 (base refresh sic3828fd; wheel superset-verified)

server.sh: pin v31 (base refresh sic3828fd; wheel superset-verified)

docker-compose.yml: pin v31 (base refresh sic3828fd; wheel superset-verified)

v30 validation: env registration (warnings 15->8) + PR#79 module (dormant)

README.md: pin v30 (EXL3 env knobs registered; sparkinfer PR#79 pcie dcp top-k module included)

server.sh: pin v30 (EXL3 env knobs registered; sparkinfer PR#79 pcie dcp top-k module included)

docker-compose.yml: pin v30 (EXL3 env knobs registered; sparkinfer PR#79 pcie dcp top-k module included)

server.sh: repair image reference to v29-gg-v20-mincaptur...@digest (previous commit corrupted the @sha256 separator)

docker-compose.yml: repair image reference to v29-gg-v20-mincaptur...@digest (previous commit corrupted the @sha256 separator)

server.sh: pin v29 image (draft-role stamp) and retire the VLLM_EXL3_TRELLIS_MIN_M workaround

docker-compose.yml: pin v29 image (draft-role stamp) and retire the VLLM_EXL3_TRELLIS_MIN_M workaround

v29: boot-without-workaround validation (draft-role stamp; #183 boot failure fixed)

Pin runtime to v29: draft-role stamp — boots with no VLLM_EXL3_TRELLIS_MIN_M workaround

needle: add fp8 KV column (18/18); 42/42 combined across both KV dtypes to ~480k tokens

v28 pin + long-context needle results (24/24 to 479k tokens)

Pin runtime to v28 (role-aware MoE runtime owner token + batch-invariant arena key)

Update tested runtime + base pin to v27 / GG-SparkInfer v20 vllm0c79e41-sie603f74

Update runtime pin to v27 (GG/SparkInfer v20 base vllm0c79e41/sie603f74)

RELEASE_TEST_SUITE: point the canonical Image section and inline presets at v26

Runtime v26: per-model runtime scoping + per-ordinal compile-cache arch key

Runtime v24: scope the EXL3 rank-sliced runtime per model (target/draft scratch isolation)

Rebase runtime onto finalized Gilded Gnosis v20 base (vLLM 5517197 / SparkInfer be0edca)

MTP layer-78 -> EXL3 tr3: model card + RELEASE_TEST_SUITE + server.sh + docker-compose (v21 image, VLLM_EXL3_TRELLIS_MIN_M=1, auto-KV)

MTP layer-78 -> EXL3-TR3 rank-sliced (malaiwah 3bpw-keep0, LDLQ-calibrated same corpus); config moe_layers [3,78], total_size -15.66GB

Add hotel-lights Max tier + independent eval (credit: malaiwah/glm52-exl3-vast) + compression context

Add independent eval harness: ORIGINAL_REPORT.md

Add independent eval harness: gpqa_diamond_summary.json

Add independent eval harness: rerun_errors.py

Add independent eval harness: gpqa_bench.py

Add independent eval harness: mathbench.py

Full release test suite: Estonia, LAVD, hotel-lights, KLD, prefill, decode C1-C8 + raw logs

Correct release test suite: Estonia 30/30 both KV dtypes, LAVD 29/30 nvfp4, KLD, power

Add release test suite report: KLD (fp8 vs nvfp4), LAVD 30x c5, Estonia harness regression

v20: rebase EXL3 backend onto Gilded Gnosis v20 canonical (vLLM 6722c1d + SparkInfer 1a88b389); MTP validated (LAVD 10/10, 0 fails, vs prior 8/10; code-gen 94-100% valid); defaults util 0.96 / blocks 1024 / max-len 262144

v20: rebase EXL3 backend onto Gilded Gnosis v20 canonical (vLLM 6722c1d + SparkInfer 1a88b389); MTP validated (LAVD 10/10, 0 fails, vs prior 8/10; code-gen 94-100% valid); defaults util 0.96 / blocks 1024 / max-len 262144

v20: rebase EXL3 backend onto Gilded Gnosis v20 canonical (vLLM 6722c1d + SparkInfer 1a88b389); MTP validated (LAVD 10/10, 0 fails, vs prior 8/10; code-gen 94-100% valid); defaults util 0.96 / blocks 1024 / max-len 262144

Ship v3 (MTP-3): pin v3 image + default MTP_TOKENS=3; update runtime + v3 validation notes

Publish validated v2 GG/Sparkinfer runtime

Add calibration + EXL3-Trellis encoder reproduction bundle (scripts + corpus + README)

Update README.md

Pin the published runtime manifest

Make the packaged server local-only by default

Clarify local and remote API addresses

Update README.md

Add Gilded Gnosis Sparkinfer Docker runtime

Update README.md

Add files using upload-large-folder tool