feat(llm-serving): serve clef-flash on candle-vllm #109

2026-11-05T01:35:56Z

rock commented

2026-11-05 01:35:56 +00:00

Serves clef-flash on candle-vllm (1 V100) and repoints the clef-serve Service at it. Service name and ClusterIP do not move, so clients are unchanged.

What

clef-candle.yaml: clef-flash, hybrid Qwen3_5 (8 attention + 24 GDN layers), 1 GPU, --dtype f16, --kv-fraction 0.30 (~66.8K KV tokens)
clef-serve.yaml: Service selector -> app: clef-candle
kustomization: clef-candle.yaml added

Why one GPU

The GDN recurrent state is per-sequence and cannot be sharded the way attention KV can.

Image

forgejo.riotpiao.com/riotpiao-poimen/candle-vllm:sm70, built by .gitea/workflows/candle-vllm-sm70.yaml on merge to main.

IMPORTANT: it must be a build containing the GDN conv fix (candle-vllm PR #5). Older :sm70 images zero the key channel band in causal_conv1d_fwd, which starves the delta rule and makes every linear-attention layer return zero. That failure is silent: the response is well formed, just wrong.

Verification vs reference

token ids 420/420 identical
beta 8647.021 vs 8646.982
delta-rule and gated-norm absmax equal to six decimals
all three head answers agree (medium / no rollback / bucket 2)

Merge order

Merge candle-vllm PR #5 and let CI publish :sm70
Merge this
Keep the reference pytorch clef-serve Deployment at replicas 0

Serves clef-flash on candle-vllm (1 V100) and repoints the clef-serve Service at it. Service name and ClusterIP do not move, so clients are unchanged. ## What - clef-candle.yaml: clef-flash, hybrid Qwen3_5 (8 attention + 24 GDN layers), 1 GPU, --dtype f16, --kv-fraction 0.30 (~66.8K KV tokens) - clef-serve.yaml: Service selector -> app: clef-candle - kustomization: clef-candle.yaml added ## Why one GPU The GDN recurrent state is per-sequence and cannot be sharded the way attention KV can. ## Image forgejo.riotpiao.com/riotpiao-poimen/candle-vllm:sm70, built by .gitea/workflows/candle-vllm-sm70.yaml on merge to main. IMPORTANT: it must be a build containing the GDN conv fix (candle-vllm PR #5). Older :sm70 images zero the key channel band in causal_conv1d_fwd, which starves the delta rule and makes every linear-attention layer return zero. That failure is silent: the response is well formed, just wrong. ## Verification vs reference - token ids 420/420 identical - beta 8647.021 vs 8646.982 - delta-rule and gated-norm absmax equal to six decimals - all three head answers agree (medium / no rollback / bucket 2) ## Merge order 1. Merge candle-vllm PR #5 and let CI publish :sm70 2. Merge this 3. Keep the reference pytorch clef-serve Deployment at replicas 0

rock added 5 commits 2026-11-05 01:35:57 +00:00

k8s(llm-serving): client-facing Service for candle-vllm coder 67760720fc

Creates coder-predictor selecting app=candle-vllm on ports 80 and 8080, so
clients keep one stable host name across the vLLM -> candle-vllm handoff and
only this selector changes from now on.

The retired KTML predictor requested two V100s and never reached Ready.

feat(llm-serving): run clef-serve so /v1/systemone answers db97159905

clef-serve was parked at 0 while candle-vllm took a card. worker-1 holds
four V100s, so the coder takes one and clef takes one, leaving two free
for a TP=2 candle run.

The gateway already routes the clef-flash model to clef-serve:80 on
/v1/systemone, so this single replica count is what turns that route on.

feat(llm-serving): run coder as fp16 14B across two V100s f6fca4705b

Replace the GPTQ W4A16 30B-A3B coder with Qwen2.5-Coder-14B-Instruct at
fp16 on TP=2.

The V100 has FP16 tensor cores and no INT4 ones, so the 4-bit path paid a
dequantisation pass on the SM. Profiling showed SM at 98% with the memory
controller at 19%, and the batch-1 step cost matched a full-model read.
Same harness, 16 streams: ITL 71.5 -> 39.7 ms, 215 -> 287 tok/s.

The deployment keeps the candle-vllm name and the app: candle-vllm label,
so coder-service.yaml and every client keep working unchanged.

fix(llm-serving): set runtimeClassName nvidia on the candle deployment 7bdc8b2de2

Without it the container runs under runc, libcuda.so is never injected and
candle-vllm panics in cudarc at startup. The pod reached Running but crashed
on every restart.

Verified live: pod Ready, coder-predictor endpoints populated, /v1/models
reports coder-14b-fp16 and a completion returns OK.

feat(llm-serving): serve clef-flash on candle-vllm 3ab127e581

Adds clef-candle.yaml: clef-flash (hybrid Qwen3_5, 8 attention + 24 GDN
layers) on one V100 via candle-vllm, and repoints the clef-serve Service
selector at it. The Service name and ClusterIP do not move, so no client
changes.

Runs on one card because the GDN recurrent state is per-sequence and cannot
be sharded the way attention KV can. --kv-fraction 0.30 gives about 66.8K KV
tokens, which is ample for a single-shot classifier-style endpoint.

The image is candle-vllm:sm70, built by .gitea/workflows/candle-vllm-sm70.yaml
on every merge to main. It must be a build containing the GDN conv fix: older
images zero the key channel band in causal_conv1d_fwd, starve the delta rule
and make every linear-attention layer return zero. That failure is silent -
the response is well formed, just wrong.

Verified against the reference implementation: token ids 420/420 identical,
beta 8647.021 vs 8646.982, delta-rule and gated-norm absmax equal to six
decimals, all three head answers agree.

rock added 1 commit 2026-11-05 01:40:40 +00:00

merge: resolve kustomization conflict, keep clef-candle and coder-14b-fp16 64a3f131b5

rock merged commit 10ce4f8109 into main

2026-11-05 01:40:43 +00:00

rock referenced this pull request

2026-11-05 01:41:37 +00:00

fix(llm-serving): hold clef-candle at replicas 0 until the image carries the fix #110

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: riotpiao-poimen/homelab#109