Issues

• Make the pull-request CI precise about which checks run, and faster (milestone 3)

  inference-sim/blis-substrate#87 · Filter by author sriumcp opened on Nov 13, 2026
  1

• Router fork CI: narrower triggers, low priority (milestone 3; the fork has issues disabled)

  inference-sim/blis-substrate#86 · Filter by author sriumcp opened on Nov 13, 2026

• Clear the release tool's refusals that the M2 substrate release (v0.14.0) carried: deprecations, milestone-1 invariant statuses

  inference-sim/blis-substrate#84 · Filter by author sriumcp opened on Nov 12, 2026

• Price the first tail layer’s outside-tail projection and KV-write work

  inference-sim/blis-substrate#82 · Filter by author namasl opened on Nov 12, 2026

• Price the fp8 indexer's native multiply rate (only the key bytes it reads change today)

  inference-sim/blis-substrate#81 · Filter by author sriumcp opened on Nov 12, 2026

• Decide whether to refit the NVFP4 grouped ramp on L17's transferred MXFP8 ramp; fix the L1/L2 documentation gap

  inference-sim/blis-substrate#80 · Filter by author sriumcp opened on Nov 12, 2026

• Attach the held-out verdicts to #66 and #67 and select (or reject) the changes by them

  inference-sim/blis-substrate#79 · Filter by author sriumcp opened on Nov 12, 2026
  1

• Replace the vendor-spec and assumed memory-size coefficients by values measured from the start-up lines of the InferenceX server logs

  inference-sim/blis-substrate#76 · Filter by author sriumcp opened on Nov 12, 2026

• DeepSeek-V4.1-Flash on H200 at 32 to 128 clients: first-token time 45% to 160% high and decode time 55% to 117% high on the sixteen rows

  inference-sim/blis-substrate#75 · Filter by author sriumcp opened on Nov 12, 2026

• Score the kernel's step times against the forward-pass measurements for MiniMax-M3, MiniMax-M2.7 and DeepSeek-V4-Pro, and use the result to settle the Blackwell first-token and EP-versus-TP errors

  inference-sim/blis-substrate#74 · Filter by author sriumcp opened on Nov 12, 2026

• DeepSeek-V4-Pro on H200 (tp8, no expert parallelism) is under-predicted 2 to 3 times in decode once the batch is above 16, and 2.3 times in first-token time at one client

  inference-sim/blis-substrate#73 · Filter by author sriumcp opened on Nov 12, 2026

• Measure the achieved fraction of line rate on the 800 Gb/s InfiniBand class instead of carrying the 400 Gb/s value over

  inference-sim/blis-substrate#72 · Filter by author sriumcp opened on Nov 12, 2026

• Measure the Rust front end's per-prompt-token admission cost on a server CPU

  inference-sim/blis-substrate#70 · Filter by author sriumcp opened on Nov 12, 2026

• Price dense MXFP8 matrix multiplies on Hopper from a transferred ramp

  inference-sim/blis-substrate#64 · Filter by author sriumcp opened on Nov 12, 2026

• Engine placement contract: place a pool through a neutral capability and refuse an engine that cannot take it

  inference-sim/blis-substrate#60 · Filter by author sriumcp opened on Nov 12, 2026

• Move engine-settings resolution into engine lowering and remove the engine-settings carrier (X2, SUB-6)

  inference-sim/blis-substrate#58 · Filter by author sriumcp opened on Nov 11, 2026

• Behavioural test cleanup

  inference-sim/blis-substrate#57 · Filter by author sriumcp opened on Nov 11, 2026

• Metric registry: total-token throughput per GPU, the P/D decode-GPU form, and interactivity as InferenceX defines them

  inference-sim/blis-substrate#52 · Filter by author sriumcp opened on Nov 11, 2026

• Network topology in the handoff price: node groups behind oversubscribed switches

  inference-sim/blis-substrate#51 · Filter by author sriumcp opened on Nov 11, 2026

• KV handoff price: base latency, transport choice and per-rail link rate instead of the all-to-all floor

  inference-sim/blis-substrate#50 · Filter by author sriumcp opened on Nov 11, 2026

• Per-KV-group transfer sizes for hybrid models have no field in plan.Transfer

  inference-sim/blis-substrate#49 · Filter by author sriumcp opened on Nov 11, 2026

• plan.Transfer for a pool that spans nodes: one From/To cannot state a move on every worker

  inference-sim/blis-substrate#48 · Filter by author sriumcp opened on Nov 11, 2026

• Router table: record the difference and a measured bound for the nearest llm-d policy, and the prefill targets in the Result

  inference-sim/blis-substrate#47 · Filter by author sriumcp opened on Nov 11, 2026

• Result and field list carry what the G1 conditions need: kernel implementation, fit range, trust flags, value classes, engine resolution

  inference-sim/blis-substrate#43 · Filter by author sriumcp opened on Nov 10, 2026

• Point record: the G2 substitution kinds the design lists and the record lacks (client_difference, later_record, router_differs)

  inference-sim/blis-substrate#41 · Filter by author sriumcp opened on Nov 10, 2026