Frank 9d7643ec5d frank-default-small: ncmoe 35->36 for prefill stability ...

Fix a live crash: cuBLAS cudaErrorInvalidValue in ggml_cuda_mul_mat_cublas
killed the service during a 145k-token prefill burst at ncmoe35/ub3072.

ncmoe 36 moves one more expert layer to CPU: frees ~0.75 GiB peak VRAM
(grid: peak 8606 vs 9360 MiB at ub3072) for -2% prefill / -2% decode.
Keeps prefill priority (top ubatch tier). Docs synced; noted fallback
GGML_CUDA_FORCE_MMQ=1 if pressure persists.

2026-08-14 10:56:40 -07:00

..

Pipeline Parallelism

docs: pipeline parallelism research + educational sample script set

2026-08-14 00:22:23 -07:00

Dankclaw Improvements.md

docs: add Dankclaw improvements research (agentic tools, runtime tuning, server/OS)

2026-08-11 18:14:44 -07:00

Dankclaw System Specs.md

docs: reflect Ling-3 vLLM (frank-default-large) on Dankclaw

2026-08-10 23:37:27 -07:00

Dankserver System Specs.md

frank-default-small: ncmoe 35->36 for prefill stability

2026-08-14 10:56:40 -07:00

deployment.md

frank-default-small: ncmoe 35->36 for prefill stability

2026-08-14 10:56:40 -07:00

DFlash DSpark Speculative Decoding.md

docs: MTP wins the DFlash replacement; record parallel premature-EOS incident

2026-08-13 23:27:06 -07:00

frank-default-large-benchmark.md

docs: MTP wins the DFlash replacement; record parallel premature-EOS incident

2026-08-13 23:27:06 -07:00

frank-default-small-benchmark.md

frank-default-small: ncmoe 35->36 for prefill stability

2026-08-14 10:56:40 -07:00

Hybrid CPU GPU Processing Dankserver.md

docs: polish pass — fix stale names, duplicates, branding, and README

2026-08-10 21:12:23 -07:00

index.md

docs: pipeline parallelism research + educational sample script set

2026-08-14 00:22:23 -07:00

ling-3.0-flash-trial.md

docs: record Ling freeze diagnosis - spec path deadlock, resolution spec-off parallel=2

2026-08-14 03:49:42 -07:00

ling-3.0-flash-vllm-plan.md

docs: add Ling-3.0-flash vLLM-ROCm deployment plan (grilled decisions, port 7770)

2026-08-10 22:02:28 -07:00

Quantization by Hardware Architecture.md

docs: reflect Ling-3 vLLM (frank-default-large) on Dankclaw

2026-08-10 23:37:27 -07:00

RAM CPU Offloading in VLLM.md

docs: polish pass — fix stale names, duplicates, branding, and README

2026-08-10 21:12:23 -07:00

Running Qwen3.6 27B FP8 on Dankclaw.md

docs: reflect Ling-3 vLLM (frank-default-large) on Dankclaw

2026-08-10 23:37:27 -07:00

Running Qwen3.6 35B A3B FP8 on Dankserver.md

docs+deploy: Dankserver now llama.cpp frank-default-small (HauhauCS Q6_K_P + DFlash)

2026-08-11 01:18:29 -07:00

Shared Memory Architecture Dankclaw.md

docs: polish pass — fix stale names, duplicates, branding, and README

2026-08-10 21:12:23 -07:00

VLLM Index.md

docs: polish pass — fix stale names, duplicates, branding, and README

2026-08-10 21:12:23 -07:00

VLLM Overview.md

docs: polish pass — fix stale names, duplicates, branding, and README

2026-08-10 21:12:23 -07:00

vllm_research_findings.md

feat: initial vLLM Qwen3.6-35B-A3B-FP8 serve script with DFlash drafting on dankserver

2026-08-07 18:28:38 -07:00