All pull requests

Pull requests

• docs: qualify four-model prefix-cache page geometry controls

  #279·gspark-etri opened Oct 24, 2026Collaborator·

  ·

• docs: record native RNGD coverage and KONI parity limits

  #275·gspark-etri opened Oct 24, 2026Collaborator·

  ·

• docs: record four-model ATOM serving validation and limits

  #272·gspark-etri opened Oct 24, 2026Collaborator·

  ·

• feat: stream RNGD KV layers into Torch and ATOM decode

  #266·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• test: evaluate held-out structured answers across four-model routes

  #264·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• exp: characterize KONI numerical drift and output-head precision limits

  #262·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• exp: validate live RNGD GPU cancellation and producer replacement

  #260·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• fix: import dense subprocess KV into hybrid GPU cache groups

  #259·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• exp: diagnose Qwen3-4B long-context RNGD/GPU output drift

  #257·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• Prefill one queued request during single-lane ATOM decode

  #255·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• Overlap GPU prefill token chunks with bounded KV copies

  #254·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• Run matched HTTP measurements on assigned ATOM and RNGD adapters

  #252·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• Measure matched four-model native, PD and data-parallel HTTP serving

  #251·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• Expose accumulated batched PD producer timings

  #250·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• fix: gate serving readiness and admission on PD producer availability

  #247·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• test: exercise four-model live PD cancellation and worker faults

  #245·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• test: validate four-model live GPU PD and numerical controls

  #243·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• feat: expose Torch PD producer cache and chunk settings

  #242·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• fix: prevent stranded prefill requests after subprocess failure

  #239·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• feat: support cache-off ATOM PD with one KV block

  #237·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• feat: connect four-model RNGD producers to live ATOM HTTP PD

  #235·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• feat: support fresh hybrid KV imports in ATOM PD scheduler

  #232·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• fix: honor tokenizer turn-end EOS in batched serving

  #230·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• exp: validate EOS-aware four-model GPU RNGD ATOM handoff

  #228·gspark-etri opened Oct 23, 2026Collaborator·

  ·

• exp: verify Gemma global causality and 8K ATOM handoff

  #226·gspark-etri opened Oct 23, 2026Collaborator·

  ·