Repository navigation

feat(qwen3): serve Qwen3-Reranker-8B on HIP behind an SGLang-compatible /v1/rerank - #1981

Merged

ashhart merged 20 commits into

ashhart:mainfrom

laudney:pr-qwen3-reranker-hip

Nov 4, 2026

Merged

feat(qwen3): serve Qwen3-Reranker-8B on HIP behind an SGLang-compatible /v1/rerank#1981
ashhart merged 20 commits into
ashhart:mainfrom
laudney:pr-qwen3-reranker-hip

Commits on Nov 4, 2026
- fix(lanes): fail queued scores once the host closes
  
  laudney
  authored and
  ashhart
  committed
  e9f7555
- feat(server): read reranker labels from 1_LogitScore
  
  laudney
  authored and
  ashhart
  committed
  6480ca0
- feat(server): serve POST /v1/rerank
  
  laudney
  authored and
  ashhart
  committed
  12e2275
- test(tools): add the Qwen3 reranker oracle
  
  laudney
  authored and
  ashhart
  committed
  61fe657
- test(server): check rerank prompt ids against transformers
  
  laudney
  authored and
  ashhart
  committed
  428ad45
- feat(qwen3): score reranker labels from the last row on the host
  
  laudney
  authored and
  ashhart
  committed
  aede63b
- test(tools): pin the reranker oracle's checkpoint and Transformers
  
  laudney
  authored and
  ashhart
  committed
  19d5a73
- feat(qwen3): serve Qwen3 rerankers on the HIP engine
  
  laudney
  authored and
  ashhart
  committed
  f0eb15f
- feat(server): accept what SGLang's /v1/rerank accepts
  
  laudney
  authored and
  ashhart
  committed
  c95cc90
- test(server): run the reranker step on every call, not from cache
  
  laudney
  authored and
  ashhart
  committed
  22b80e5
- feat(server): answer /v1/rerank in #1895's agreed shape
  
  laudney
  authored and
  ashhart
  committed
  60c626c
- refactor(rerank): declare the reranker in the engine's Info
  
  laudney
  authored and
  ashhart
  committed
  334a6fa
- docs(rerank): document /v1/rerank and the score queue's drain
  
  laudney
  authored and
  ashhart
  committed
  6f42d81
- test(rerank): an FP32 reference and the model card's formatter check
  
  laudney
  authored and
  ashhart
  committed
  7649399
- test(rerank): gate the scores against the FP32 reference
  
  laudney
  authored and
  ashhart
  committed
  122e3e3
- feat(rerank): cut content to max_length and use the server envelope
  
  laudney
  authored and
  ashhart
  committed
  1033a8c
- docs(rerank): document the max_length cut and the error envelope
  
  laudney
  authored and
  ashhart
  committed
  9848915
- build(rerank): move the rerank host step into its own build file
  
  laudney
  authored and
  ashhart
  committed
  c79e7ba
- docs(recipes): serve Qwen3-Reranker-8B on native HIP
  
  laudney
  authored and
  ashhart
  committed
  c4a8766
- build: cuda.zig stays at 600 lines with the rerank steps (one-use eos_family inlined)
  ashhart
  committed
  70489e9

Commits

Commits on Nov 4, 2026