test(frank-medium): A/B — flip to DFlash2 drafter (SPEC_METHOD=dflash, SPEC=7)
This commit is contained in:
@@ -110,11 +110,12 @@ Environment=MAXLEN=262144
|
|||||||
# 400 here. Vision traffic routes to frank-small on dankclaw (harness
|
# 400 here. Vision traffic routes to frank-small on dankclaw (harness
|
||||||
# modelRoles vision -> frank-small).
|
# modelRoles vision -> frank-small).
|
||||||
Environment=GPU_UTIL=0.98
|
Environment=GPU_UTIL=0.98
|
||||||
# Speculative decoding: MTP self-speculation at depth 8 (the model's own
|
# Speculative decoding: A/B test — DFlash2 external drafter at depth 7 (the
|
||||||
# single-layer MTP head, applied autoregressively). The documented plain-FP8
|
# documented DFlash2 depth), to compare against the MTP self-speculation
|
||||||
# path (upstream docker-compose) uses num_speculative_tokens=8.
|
# baseline (depth 8, mean accept 5.66/8). The DFlash2 drafter is a separate
|
||||||
Environment=SPEC_METHOD=mtp
|
# quantization-agnostic checkpoint at /models/Qwen3.8-27B-DFlash2-FP8.
|
||||||
Environment=SPEC=8
|
Environment=SPEC_METHOD=dflash
|
||||||
|
Environment=SPEC=7
|
||||||
# KV cache: 0 = profile the shape (the documented plain-FP8 path's default;
|
# KV cache: 0 = profile the shape (the documented plain-FP8 path's default;
|
||||||
# the upstream docker-compose leaves the pin commented out). The measured pin
|
# the upstream docker-compose leaves the pin commented out). The measured pin
|
||||||
# for this shape is 15989735424 (14.89 GiB, upstream, 2x R9700 32GB); after
|
# for this shape is 15989735424 (14.89 GiB, upstream, 2x R9700 32GB); after
|
||||||
|
|||||||
Reference in New Issue
Block a user