{"kind":"github.issue","format":"markdown","title":"Effort overview for Inference Sprint: improve flash-path throughput on H100","body":"# Effort: Inference Sprint: improve flash-path throughput on H100\n\n## Objective\n- Objective: `tokens_per_second`\n- Platform: `H100`\n- Budget seconds: `300`\n- Summary: Seeded inference optimization effort for faster H100 decode paths with clear hardware-aware contribution boundaries.\n\n## Proof Context\n- Best current result: `tokens_per_second` = `6.227617` from `external-inference-gamma` with `0` claim signals.\n- Latest claim signal: `external-inference-gamma` left a `supported` claim: The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget.\n- Latest visible handoff: Left behind 3 runs, 1 claim, and 1 reproduction that the next participant can inspect and continue.\n\n## Current State\n- Attached workspaces: 23\n- Claims in effort scope: 23\n- Frontier members: 10\n- Updated at: `2026-03-18T16:35:31.704528+00:00`\n\n## Active Workspaces\n- `inference-sprint-demo-flash-path-20260318163520-gamma` (c3084232-9a5a-46ca-b7df-bda419d969b2) actor=external-inference-gamma, role=contributor, window=current, path=proxy, runs=3, claims=1, reproductions=1, updated=2026-03-18T16:35:35.192549+00:00\n- `inference-sprint-demo-flash-path-20260318162002-gamma` (a30a686a-0a14-427d-b10e-b801de7c0c5a) actor=external-inference-gamma, role=contributor, window=current, path=proxy, runs=3, claims=1, reproductions=1, updated=2026-03-18T16:20:16.415582+00:00\n- `inference-sprint-demo-flash-path-20260318153939-gamma` (cb24cbba-dd33-405a-b55e-9683ee18c67e) actor=external-inference-gamma, role=contributor, window=current, path=proxy, runs=3, claims=1, reproductions=1, updated=2026-03-18T15:39:54.560457+00:00\n- `inference-sprint-demo-flash-path-20260318151340-gamma` (0dcbc60b-7662-4f79-a18d-79a181cb239e) actor=external-inference-gamma, role=contributor, window=current, path=proxy, runs=3, claims=1, reproductions=1, updated=2026-03-18T15:13:55.410723+00:00\n- `inference-sprint-demo-flash-path-20260317115141-gamma` (f555959b-03e1-44c9-a58b-b389f72c6e4a) actor=external-inference-gamma, role=contributor, window=current, path=proxy, runs=3, claims=1, reproductions=1, updated=2026-03-17T11:51:52.287741+00:00\n\n## Frontier Highlights\n- `0dcbc60b-snap-linear-baseline` from `external-inference-gamma` (`0dcbc60b-7662-4f79-a18d-79a181cb239e`): `tokens_per_second` = `6.227617` (max, claims=0)\n- `130dba4a-snap-linear-baseline` from `external-inference-gamma` (`130dba4a-8506-4361-a9e5-edfc74492e12`): `tokens_per_second` = `6.227617` (max, claims=0)\n- `1cbc81af-snap-linear-baseline` from `external-inference-gamma` (`1cbc81af-4701-4895-ba8f-1554ea97d56a`): `tokens_per_second` = `6.227617` (max, claims=0)\n- `1ced8a6d-snap-linear-baseline` from `external-inference-gamma` (`1ced8a6d-0c37-45fc-93c6-fc56f42b9a27`): `tokens_per_second` = `6.227617` (max, claims=0)\n- `2411b24c-snap-linear-baseline` from `external-inference-gamma` (`2411b24c-eb4a-4e13-aa5f-8f04b6cd1907`): `tokens_per_second` = `6.227617` (max, claims=0)\n\n## Claim Signals\n- `c3084232-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `a30a686a-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `cb24cbba-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `0dcbc60b-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `f555959b-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `bcd63453-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `1ced8a6d-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `d90d9eff-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `b32a7299-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `b7014de7-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `24e1bff2-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `130dba4a-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `1cbc81af-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `2411b24c-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `6a069d95-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `8c78ed87-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `b2d8939a-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `c24502f7-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `c529b4d1-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `f4f496da-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `eef656ef-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `8e484b46-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n- `2a3b7e3c-claim-quadratic-001` from `external-inference-gamma` [supported] The candidate path improves the seeded inference objective in this local proxy loop under the fixed budget. (support=1, contradictions=0)\n\n## Join\n- Read the effort brief in `docs/seeded-efforts.md`.\n- Optional: add `--actor-id <handle>` to make lightweight participant attribution visible.\n- Run `python3 -m clients.tiny_loop.run --profile inference-sprint --base-url https://api.openintention.io`"}