Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 0 additions & 1 deletion docs/reference-matrix.md
Original file line number Diff line number Diff line change
Expand Up @@ -63,7 +63,6 @@ Reading rules:
| NVIDIA/TensorRT-LLM @ f33e0e5fd | skill-doc-only (`.claude/skills/`) | mined — evidence taxonomy, kernel writing, casebook, roofline refs → `dense-performance-evidence`, `kernel-writing-guide`, `inference-optimization-casebook`, `performance-ceiling-analysis` | mined — host overhead verdict + line-profiling loop → `host-overhead-profiling`, `host-overhead-optimization` | mined — weight-movement/overlap + KV host-cache patterns → `weight-loading-optimization`; disk-level: negative (see below) | out-of-scope — host/device weight-movement overlap remains in RAM/Storage, not network coverage | mined — CPU affinity rule + env preflight → `os-process-placement` |
| Ascend/MindIE-SD @ 8481d39 | skill-doc-only (`.agents/skills/`) | mined — 5-layer analysis, bottleneck casebook → `ascend-performance-analysis`, `ascend-optimization-casebook` | mined — host-opt async/pattern sections → `weight-loading-optimization` | mined — load-path judgment, fixed-overhead accounting, offload tiers, weights-prep → `weight-loading-optimization` | mined — USP strategy, masking formula, topology rules → `ascend-distributed-parallel` | mined — device health, card-group contracts, measurement discipline → `ascend-device-health` |
| ascend-ai-coding/awesome-ascend-skills @ 75ec89c | Historical skill/document import (`skills/`, `external/cannbot/`), not whole native repository mining; separate communication-wrapper source slice pinned to `75ec89ce9d401be8ac95980e1c3f42f8d4315057` | mined — graph mode, operator profiling/tuning, tiling design → `ascend-graph-mode`, `ascend-operator-profiling`, `ascend-operator-tuning`, `ascendc-tiling-design` | assessed/deferred — KV overview and async-stream recipe lack a complete transfer/completion implementation and retain cannbot license UNKNOWN; [boundary/readset](reviews/issue91-awesome-ascend.md#host-kv-boundary); no new RAM capability | mined — device-side prefetch + diffusers CPU-offload playbook → `weight-loading-optimization`; disk-level: negative | mined (wrapper-source only) — native/PyTorch test selection, tensor-byte meaning, timing/oracle coverage and rank-map counterexamples → `ascend-distributed-parallel`; [full-pin readset](../src/inferrsi/knowledge/data/skill_library/hardware/ascend/ascend-distributed-parallel/references/awesome-ascend-communication-tests.md); no native-collective or transport audit | mined — env checks, device-busy preflight, thread/env tables → `os-process-placement` |
| cann/cannbot @ `c0e2ad23ee1a73e3c0f84c8df1cc459e2aae3110` (selected official plugin: CANN-2.0) | Direct canonical Git source inspection, 2026-10-23; fixed-stage model-inference workflow, three role instructions, registered callbacks and archive helper; [readset/trace/license](../src/inferrsi/knowledge/data/skill_library/hardware/ascend/ascend-performance-analysis/references/cannbot-stage-evidence.md); dependent skills/recipes and host execution unread/unrun | mined (one source-derived rule) — current-stage execution, actual quality result and comparable baseline versus lexical callback completion → [ascend-performance-analysis](../src/inferrsi/knowledge/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md); no numerical/backend/performance qualification | out-of-scope — no host-memory mechanism extracted | out-of-scope — progress-file I/O is not storage optimization | out-of-scope — workflow instructions do not establish collective/transport implementation | out-of-scope — agent workflow is not OS scheduling |
| vllm-project/vllm-ascend @ bc8115ed | skill-doc-only (`.agents/`); [version/responsibility routing](../src/inferrsi/knowledge/data/skill_library/README.md#vllm-ascend-version-and-responsibility-routing) | mined — FP8/KV-replication/multimodal adaptation → `ascend-model-adaptation` | unknown — runtime host logic not in sparse checkout | mined — FP8 paired-tensor loading pattern → `ascend-model-adaptation` (Configure cell); runtime loader not in checkout | unknown | mined — Dockerfile thread/env ship defaults → `os-process-placement` (historical defaults only) |
| vllm-project/vllm @ `7f730b4cea9a66520a9200265928ebe016b48a39` (Apache-2.0) | Dot bundle-mediated; `scheduler.py` admission/budgets/preemption/deferred frees and connector `v1/base.py` contracts/consumers; concrete connector, worker, cache-manager internals, config schema and tests unread | mined — independent admission budgets and async-KV readiness → `diagnose-vllm-admission-pressure`; completion/retention/failure ownership → `audit-vllm-kv-lifetimes`; conditional proposals and unrun verification only | out-of-scope — device KV is not host RAM | out-of-scope — external-hit API establishes no physical tier | out-of-scope — connector contract is not transport implementation | out-of-scope — engine scheduling is not OS priority control |
| vllm-project/vllm-ascend @ `8a8e1e05ecb64a30804c99c8f362909f9a0cb5bf` (Apache-2.0; separate runtime pin) | Dot bundle-mediated; `compilation/acl_graph.py`, pinned graph-mode guide and selected `test_acl_graph.py` cases; tests inspected by supplier, not run; caller/ENPU ordering and helpers untraced; native single-tier spec/worker, registration and DMA binding read directly ([seven-file ledger](../src/inferrsi/knowledge/data/skill_library/hardware/host/host-memory-management/references/vllm-ascend-native-offload-readset.json)); external vLLM dependencies unpinned; [separate model-adapter routing](../src/inferrsi/knowledge/data/skill_library/README.md#vllm-ascend-version-and-responsibility-routing) | mined — effective dispatch, caller-owned stable inputs, branch-specific replay/update order and capture-resource failures → `ascend-aclgraph-capture-replay`; separate from the older skill-doc pin | mined — normal event-query-gated CPU descriptor-pool ownership and wait/recycle distinction → [host-memory-management](../src/inferrsi/knowledge/data/skill_library/hardware/host/host-memory-management/SKILL.md); source candidate, native tests not run; shared/tiering and allocator behavior excluded | out-of-scope | out-of-scope — offloader synchronization is not network evidence | out-of-scope — ACLGraph runtime supplies no new OS-control evidence; older Dockerfile/thread-env defaults remain scoped to bc8115ed |
Expand Down
81 changes: 0 additions & 81 deletions docs/reviews/issue91-cannbot.md

This file was deleted.

1 change: 0 additions & 1 deletion docs/searchspace-matrix.md
Original file line number Diff line number Diff line change
Expand Up @@ -49,7 +49,6 @@ The primary compute and HBM resource. Most current work lives here.
| **Optimize** | Bounded experiment selection, distinct mechanism hypotheses, incumbent preservation and final-artifact integrity | `optimization-search-policy` | [Sarathi chunk-budget case](../src/inferrsi/knowledge/data/skill_library/methods/optimization-search-policy/cases/sarathi-chunk-budget.md): decode-first accounting, prefill overhead and TTFT/TBT tradeoffs; no universal budget. [Infera provenance](../src/inferrsi/knowledge/data/skill_library/methods/calibrate-serving-simulation/references/infera-serving-search/README.md) adds a seed-only comparator and evaluated-trial ledger example; its trace-ranking check is proposed. Methodology provenance stays in Reference Matrix §2; no hardware coverage from agent-paper scores | vllm-evolve @ 43cc82c; scoped A/F paper amendments; separate Sarathi case and bounded Infera Git readset in Reference Matrix | 🟢 (source-derived candidate procedure; empirical not_run) |
| **Optimize** | Bounded multistream overlap: dependency-DAG constrained (data/state/event/collective_order edges; shared input ≠ dependency), minimal-first, execution-mode sync contract (eager streams vs TorchAir) | `cann-attribution-overlap` | Metric separation (wall / union-of-busy / summed durations / summed duration+wait; repeated-instance product = estimate); device-memory domain qualification; formula conflicts carried unresolved | cann/oam-tools @ 8440935f (pack-mediated, wave 6) | 🟢 |
| **Verify** | Evidence taxonomy (L0 correctness → L5 production), claim-evidence matching | `dense-performance-evidence`, `rsi-evidence`, `claim-contract` | InferRSI PASS/FAIL/INCONCLUSIVE protocol | InferRSI @ 4c8e1285 | 🟢 |
| **Diagnose / Verify** | Fixed-stage PyTorch/Ascend model report: bind current stage and actually loaded candidate to quality-validator output and comparable baseline | [ascend-performance-analysis](../src/inferrsi/knowledge/data/skill_library/hardware/ascend/ascend-performance-analysis/SKILL.md) | [Stage-evidence rule](../src/inferrsi/knowledge/data/skill_library/hardware/ascend/ascend-performance-analysis/references/cannbot-stage-evidence.md): workflow instructions versus registered marker/keyword checks; completion is not numerical validation | canonical cann/cannbot @ c0e2ad23ee1a73e3c0f84c8df1cc459e2aae3110; selected CANN-2.0 scope | Candidate procedure only; host/model/backend not_run, agent efficacy not evaluated; no new measured coverage |
| **Verify** | Same-window A/B protocol, noise floor, acceptance-state discipline; correctness-first overlap acceptance (fallback works, no KV/state reorder, actual overlap, no outside-module regression) | `ascend-optimization-casebook` (evidence-required section per case), `dense-performance-evidence`, `rsi-evidence`, `cann-attribution-overlap` | MindIE-SD perf-gate protocol: exploration vs acceptance state, three-proof requirement | MindIE-SD perf-gate @ 8481d39b; cann/oam-tools @ 8440935f | 🟢 |
| **Observe / Configure / Optimize** | DALI CPU/GPU queue overlap and operator buffer sizing under explicit executor-mode gates | `training-input-pipeline` | Separated queues require async+pipelined and non-dynamic mode; output synchronization retained | Reference B1, DALI @ 6fe47c48d3b0 | 🟢 (Python API evidence, not C++/device validation) |
| **Diagnose / Configure / Optimize** | Training capacity versus throughput, parameter/optimizer tier budgets, recomputation/offload choices; DCP exposed staging pause | `training-resource-budget`, `dcp-async-save-performance` | DeepSpeed stage gates; expert MFU/HFU denominator discipline; historical MindSpeed selection/allocator/rescue source boundaries | Reference CR1, CR6, CR5, C1 async-only | 🟢 (MindSpeed historical source candidate; current support unqualified) |
Expand Down
Loading
Loading