-
bf64c0d9c2
perf: frank-default-large batch 12288 / ubatch 4096
Frank
2026-08-11 16:36:39 -07:00
-
f5a959bfa2
fix: disable --fit in frank-default-large serve (full offload fits; fit aborted on explicit n-gpu-layers)
Frank
2026-08-11 16:28:03 -07:00
-
25e90bbbf7
deploy: frank-default-large final config — ROCm + draft-dflash, 3x200700 ctx (bench-fastest)
Frank
2026-08-11 16:27:15 -07:00
-
d0c80b5b1a
bench: 4k pipeline harness w/ early crash check; parametrized spec-type
Frank
2026-08-11 16:14:44 -07:00
-
df185c8d22
feat: full-pipeline 64k spec benchmark harness
Frank
2026-08-11 15:07:21 -07:00
-
aa2f4bfe39
feat: frank-default-large → llama.cpp Qwen3.6-27B Q8_K_P + DFlash draft + ngram-mod (bench pending)
Frank
2026-08-11 13:47:37 -07:00
-
a74c1c5bcb
Lower dankserver compute threads 8→7 to leave CPU headroom
Frank
2026-08-11 13:36:36 -07:00
-
df41aaafe7
Offload one more MoE expert to CPU on frank-default-small (32→33 of 40)
Frank
2026-08-11 13:13:17 -07:00
-
4c5a4a5f3b
docs: frank-default-small — 4 parallel slots, batch 8192/2048, n-cpu-moe 32
Frank
2026-08-11 03:13:36 -07:00
-
870376d75f
feat: frank-default-small — 4 slots, batch 8192/2048, n-cpu-moe 32, ngram-mod
Frank
2026-08-11 02:54:49 -07:00
-
0aabd57c69
docs: README — frank-default-small runtime → ngram-mod (DFlash retired)
Frank
2026-08-11 02:42:32 -07:00
-
677b4af2a4
docs: frank-default-small now uses ngram-mod, DFlash retired
Frank
2026-08-11 02:41:33 -07:00
-
fd528ab364
chore: frank-default-small — ngram-mod, drop DFlash draft
Frank
2026-08-11 02:29:32 -07:00
-
5c9e812221
docs: frank-default-small now 200000 ctx/stream, batch 3072/1536, --no-mmap
Frank
2026-08-11 01:52:12 -07:00
-
9971f83e8b
perf: frank-default-small ctx 200000/stream, --no-mmap, batch 3072/1536
Frank
2026-08-11 01:45:41 -07:00
-
e45f457ad4
perf: frank-default-small batch 3072 / ubatch 1536 + ctx-checkpoints
Frank
2026-08-11 01:33:28 -07:00
-
35c44c4123
docs+deploy: Dankserver now llama.cpp frank-default-small (HauhauCS Q6_K_P + DFlash)
Frank
2026-08-11 01:18:29 -07:00
-
6a86a9df93
fix: frank-default-small ctx-size 786432 (3 slots x 262144 each)
Frank
2026-08-11 01:05:24 -07:00
-
f93096f2bf
feat: frank-default-small = HauhauCS Qwen3.6-35B-A3B uncensored Q6_K_P + DFlash on dankserver
Frank
2026-08-11 00:57:04 -07:00
-
903d54c389
docs: reflect Ling-3 vLLM (frank-default-large) on Dankclaw
Frank
2026-08-10 23:37:27 -07:00
-
52ad985978
feat: raise Ling-3 GPU memory utilization to 0.95
Frank
2026-08-10 23:25:02 -07:00
-
bf14ac7a7b
fix: run Ling-3 vLLM from a cwd with no top-level vllm/ dir
Frank
2026-08-10 23:16:46 -07:00
-
4a17948443
feat: add Ling-3.0-flash vLLM service for Dankclaw (frank-default-large, port 7770)
Frank
2026-08-10 22:36:37 -07:00
-
91188d4b1b
docs: add Ling-3.0-flash vLLM-ROCm deployment plan (grilled decisions, port 7770)
Frank
2026-08-10 22:02:28 -07:00
-
2a7571abdc
chore: remove stale root-level duplicates of scripts/ files
Frank
2026-08-10 21:12:48 -07:00
-
f650fff0aa
docs: polish pass — fix stale names, duplicates, branding, and README
Frank
2026-08-10 21:12:23 -07:00
-
993b6af576
docs: add DFlash/DSpark speculative decoding research + quantization-by-hardware guide
Frank
2026-08-10 21:09:17 -07:00
-
90c4883f89
feat: add consistent deployment/update process (deploy.sh + docs + AGENTS.md)
Frank
2026-08-10 21:04:37 -07:00
-
3a64c8f3dd
docs: fact-check knowledge base against live machines and vLLM source
Frank
2026-08-10 20:59:33 -07:00
-
841d872aa7
sync scripts from remotes: ROCmFPX llama-server.service (dankclaw), dankserver vLLM serve/validate; drop failed-attempt scripts
Frank
2026-08-10 19:48:46 -07:00
-
ced4c6f46e
sync: update vllm-awq-qwen36-dflash-ampere.sh from dankserver
Frank
2026-08-10 03:23:09 -07:00
-
11d66c3823
Set max-model-len to 32768 for testing
Frank
2026-08-09 22:19:32 -07:00
-
6bdabfc65b
Add VLLM_USE_V2_MODEL_RUNNER=0 to skip V2 model runner MARLIN repacking OOM
Frank
2026-08-09 21:49:25 -07:00
-
83b0bea29b
Fix: move commented speculative config to end to not break bash line continuation
Frank
2026-08-09 12:27:28 -07:00
-
88aa18ef34
Reduce GPU memory to 75%, add expandable_segments to fix OOM
Frank
2026-08-09 12:11:48 -07:00
-
9dfa16fae9
Disable speculative decoding - OOM on stable vLLM too
Frank
2026-08-09 12:08:26 -07:00
-
578c3e8f71
Enable DFlash speculative decoding (12 tokens), increase CPU offload to 20GB
Frank
2026-08-09 11:54:40 -07:00
-
8c13c7bbaa
Add AGENTS.md git workflow rules and initial vllm-awq-qwen36-dflash-ampere.sh from dankserver
Frank
2026-08-09 11:54:04 -07:00
-
ad96733340
Merge PR #1: Update AGENTS.md with target-specific branching and service status
Frank
2026-08-09 11:03:12 -07:00
-
-
1f85e57863
Update AGENTS.md with target-specific branching and service status
vllm-qwen36
Frank
2026-08-09 10:37:57 -07:00
-
6bddf0a72a
Add AGENTS.md, deploy script, update README with service status and file locations
Frank
2026-08-09 09:47:24 -07:00
-
469bfaf979
Update README with validation instructions and performance data
Frank
2026-08-08 21:46:00 -07:00
-
726b9a35ec
Add validator script with target selection and test suite
Frank
2026-08-08 21:42:37 -07:00
-
5ed9bf0410
Update documentation with AWQ+DFlash lessons learned and improved validation script
Frank
2026-08-08 21:29:23 -07:00
-
c530e16880
Temporarily disable speculative decoding to test base model
Frank
2026-08-08 17:28:00 -07:00
-
9d00f36d2b
Lower context to 32k, fix duplicate lines, remove prefix caching
Frank
2026-08-08 17:03:16 -07:00
-
9227e1a8b9
Add AWQ DFlash Ampere script and improved validation
Frank
2026-08-08 16:41:44 -07:00
-
659fba6a5d
Lower --max-num-batched-tokens to 1536 to free VRAM for weights
Frank
2026-08-08 12:56:28 -07:00
-
9c1f777423
Fix serve-vllm-qwen36-dflash.sh: remove 'vllm serve' prefix (image entrypoint handles it)
Frank
2026-08-08 12:15:31 -07:00
-
4edcada2a8
Increase max-model-len to 32768, cpu-offload to 40GB, add reasoning-parser and tool-call-parser
Frank
2026-08-08 12:03:23 -07:00
-
-
-
69bb41057c
Add --enable-auto-tool-choice and --tool-call-parser qwen3_xml for tool choice support
Frank
2026-08-08 11:23:57 -07:00
-
dd19627635
Update validation script to use frank-default-small model name
Frank
2026-08-08 10:38:30 -07:00
-
35bd22cd75
Add --served-model-name frank-default-small
Frank
2026-08-08 10:35:24 -07:00
-
0f97dcf449
Update validation script to test parallel streams (4 concurrent requests)
Frank
2026-08-08 00:30:08 -07:00
-
6a7025d01e
Iteration 2: Remove --enforce-eager to enable torch.compile (+14.6% prompt speed: 5.95 -> 6.82 tok/s)
Frank
2026-08-07 23:07:51 -07:00
-
1b90f421c6
Initial vLLM Qwen3.6-35B-A3B-FP8 serving with CPU expert offloading
Frank
2026-08-07 22:20:52 -07:00
-
9971bdef0f
feat: initial vLLM Qwen3.6-35B-A3B-FP8 serve script with DFlash drafting on dankserver
Frank
2026-08-07 18:28:38 -07:00
-
5452bb63f1
Add vLLM serve script, systemd service, and validation test for Qwen3.6-35B-A3B-FP8 on dankserver
Frank
2026-08-07 18:19:40 -07:00
-
663e47866b
Bump context to 600000 (200000/stream x 3 parallel)
Frank
2026-08-07 17:18:10 -07:00
-
a5f72627bb
Use --alias frank-default-large, multi-stream validation script
Frank
2026-08-07 17:15:09 -07:00
-
471e963350
Fix parallel validation, remove ngram-mod
Frank
2026-08-07 16:44:12 -07:00
-
2426ca563a
Add ngram-mod to spec pipeline: -1.2 tok/s regression
Frank
2026-08-07 16:25:49 -07:00
-
764a65080f
Validation: 2 parallel prompts, ~18 tok/s with DFlash
Frank
2026-08-07 16:21:20 -07:00
-
25d57184a6
Enable DFlash speculative decoding, remove mlock
Frank
2026-08-07 16:10:55 -07:00
-
1949f0651b
Pivot to llama.cpp with DFlash speculative decoding on ROCm
Frank
2026-08-07 15:41:27 -07:00
-
2bb1b2ae1d
Baseline: ThinkingCap 27B FP8 on dankclaw with enforce-eager and validation timeout
Frank
2026-08-07 08:58:20 -07:00
-
e4fba35e89
Baseline: ThinkingCap 27B FP8 on dankclaw
Frank
2026-08-07 08:49:59 -07:00
-
7dce375ab4
docs: fix dankclaw specs (30GiB RAM, 96GiB GPU-visible, ROCm 7.2.4) + add validate.sh
Frank
2026-08-06 22:40:44 -07:00
-
5076ef828f
chore: fix typo in dankclaw specs, format index table
Frank
2026-08-06 22:29:23 -07:00
-
dc356acce7
docs: add top-level knowledge base index
Frank
2026-08-06 00:16:03 -07:00
-
671129ea69
chore: ignore obsidian workspace files
Frank
2026-08-06 00:14:06 -07:00
-
39db415348
feat: initialize vLLM scripts knowledge base
Frank
2026-08-05 23:49:02 -07:00