tt-stack

1-card target: ThinkingCap-27B on ONE P150a (TP=1) - context ladder, accuracy, bf4 coverage, batching

3

#69 opened 2026-09-25 10:06:18 +02:00 by Grok

Quantize on TT cards: CUDA-free GPTQ/LDLQ bfp4 pipeline (Hessians, GPTQ sweep, FP8 teacher on Blackhole)

3

#67 opened 2026-09-25 08:39:50 +02:00 by Grok

DSpark drafter (ThatchCloud) on ThinkingCap

3

#66 opened 2026-09-24 21:54:47 +02:00 by Grok 2 / 5

ThinkingCap FP8 baseline on TT

#65 opened 2026-09-24 14:59:27 +02:00 by Grok

upstream scout 2026-09-24: tt-metal/ttnn/vllm-tt perf changes relevant to Qwen3.8 on 2xP150 (top-5 + pin-bump hazards)

2

#64 opened 2026-09-24 14:22:12 +02:00 by Grok

MTP on ThinkingCap-Qwen3.8-27B-FP8 (TP=2): inventory, per-row B=4/8 design, go/no-go, phased gates

19

#63 opened 2026-09-24 11:41:26 +02:00 by Grok 1 / 8

bfp-quant: EXL3-style Hessian-aware offline quantizer → native bfp4_b (grid-snapped), + weight accuracy gate

8

#62 opened 2026-09-24 11:11:53 +02:00 by Grok 0 / 5

research: exl3xpu / EXL3 trellis quant on Blackhole — port is no-go; EXL3-style offline bf4 quantizer + accuracy gate instead

1

#61 opened 2026-09-24 10:46:54 +02:00 by Grok

research: Homa transport for Blackhole card networking — viability (verdict: no-go)

#60 opened 2026-09-24 10:45:17 +02:00 by Grok

hyp F: DRAM prefetcher sub-device is the only in-tree way to hide the residue under the weight read

1

#58 opened 2026-09-23 23:01:11 +02:00 by Grok

hyp E: TP collectives, not the second card's DRAM, are a slice of the 48.6 ms step

4

#57 opened 2026-09-23 21:34:27 +02:00 by Grok

hyp D: non-matmul decode ops share the matmul Tensix grid, so they cannot hide under the weight read

4

#56 opened 2026-09-23 21:34:27 +02:00 by Grok

hyp C: ~15 ms of non-matmul device work sits inside the traced ttnn call

12

#55 opened 2026-09-23 21:34:27 +02:00 by Grok

hyp B: 8.2 ms matmul floor is serial with the weight read (necessary, not sufficient)

3

#54 opened 2026-09-23 21:34:26 +02:00 by Grok

P150 single-stream decode is 48.6 ms; gate is ≤25.0 ms, no MTP

4

#53 opened 2026-09-23 21:34:26 +02:00 by Grok

serve: make Phase 2a (device-resident decode) & live prefix caching durable in NixOS module

1

#52 opened 2026-09-23 15:38:13 +02:00 by multica-agent 0 / 4

prefill: widen MLP gate|up AGMM and GDN in-proj grids from 8x9 to 11x10 full Blackhole mesh

3

#51 opened 2026-09-23 15:38:01 +02:00 by multica-agent 0 / 4

prefill: capture and replay masked-bucket traces (<=2048 tok) to eliminate ~420ms host dispatch floor

3

#50 opened 2026-09-23 15:36:47 +02:00 by multica-agent 0 / 4

ops: production tool calling, the hugepage reduction and SDPA streaming are all /run-only and do not survive a reboot

2

#49 opened 2026-09-20 08:55:58 +02:00 by Grok

prefix caching: deploy decision — the correctness gate is clear, this is the largest remaining makespan win

2

#48 opened 2026-09-20 08:55:31 +02:00 by Grok