ViperEkura · 0z5a · Sep 2, 2026 · Sep 2, 2026
diff --git a/astrai/extension/backend/linear.py b/astrai/extension/backend/linear.py
@@ -1,9 +1,9 @@
 """Inference-only dispatch for AstrAI linear layers.
 
 The CUDA GEMV path is deliberately narrow: automatic selection is enabled
-only for single-row BF16 shapes measured to beat ``F.linear`` on a supported
-architecture. Every training, prefill, unsupported-layout, and unmeasured
-call falls back to PyTorch.
+only for small decode batches and BF16 shapes measured to beat ``F.linear``
+on a supported architecture. Every training, prefill, unsupported-layout,
+and unmeasured call falls back to PyTorch.
 """
 
 import logging
@@ -30,20 +30,72 @@
 logger = logging.getLogger(__name__)
 
 # Shape keys are (N, K) for Y[M, N] = X[M, K] @ W[N, K].T. A band is
-# automatic only after both the per-shape >=5% and end-to-end decode >=3%
-# gates pass and checkpoint greedy output remains stable. M=1 and M=8 remain
-# empty on SM89; the safe M=2/4 bands improve real-engine throughput by
-# 11.8-14.0%.
+# automatic only after both the per-shape >=5% and projection-chain/engine
+# >=3% gates pass and output argmax remains stable. M=1 is limited to OPT 1.3B;
+# M=8 remains empty because at least one projection in each measured family
+# misses the per-shape gate even when its aggregate chain result is positive.
+_COMMON_TRANSFORMER_SM89_SHAPES = frozenset(
+    {
+        (1024, 4096),  # LLaMA 3 8B K/V
+        (4096, 4096),  # LLaMA 2/3 7B/8B Q/O
+        (11008, 4096),  # LLaMA 2 7B gate/up
+        (4096, 11008),  # LLaMA 2 7B down
+        (14336, 4096),  # LLaMA 3 8B gate/up
+        (4096, 14336),  # LLaMA 3 8B down
+        (5120, 5120),  # LLaMA 2 13B Q/K/V/O
+        (13824, 5120),  # LLaMA 2 13B gate/up
+        (5120, 13824),  # LLaMA 2 13B down
+        (16384, 4096),  # GPT-NeoX MLP up
+        (4096, 16384),  # GPT-NeoX MLP down
+    }
+)
+_COMMON_TRANSFORMER_SM89_M4_SHAPES = _COMMON_TRANSFORMER_SM89_SHAPES - {
+    (4096, 4096),
+    (11008, 4096),
+    (4096, 11008),
+}
+_QWEN2_7B_SM89_SHAPES = frozenset(
+    {
+        (512, 3584),  # K/V
+        (3584, 3584),  # Q/O
+        (18944, 3584),  # gate/up
+        (3584, 18944),  # down
+    }
+)
+_LLAMA3_70B_SM89_SHAPES = frozenset(
+    {
+        (1024, 8192),  # K/V
+        (8192, 8192),  # Q/O
+        (28672, 8192),  # gate/up
+        (8192, 28672),  # down
+    }
+)
+_OPT_1_3B_SM89_SHAPES = frozenset(
+    {
+        (2048, 2048),  # Q/K/V/O
+        (8192, 2048),  # MLP up
+        (2048, 8192),  # MLP down
+    }
+)
+
 _AUTO_GEMV_SHAPES: dict[tuple[int, int], dict[int, frozenset[tuple[int, int]]]] = {
     (8, 9): {
-        2: frozenset(
+        1: _OPT_1_3B_SM89_SHAPES,
+        2: _COMMON_TRANSFORMER_SM89_SHAPES
+        | _QWEN2_7B_SM89_SHAPES
+        | _LLAMA3_70B_SM89_SHAPES
+        | _OPT_1_3B_SM89_SHAPES
+        | frozenset(
             {
                 (256, 1536),
                 (1536, 1536),
                 (100000, 1536),
             }
         ),
-        4: frozenset({(256, 1536), (1536, 1536)}),
+        4: _COMMON_TRANSFORMER_SM89_M4_SHAPES
+        | _QWEN2_7B_SM89_SHAPES
+        | _LLAMA3_70B_SM89_SHAPES
+        | frozenset({(256, 1536), (1536, 1536)}),
     }
 }
 _AUTO_GEMV_M = frozenset(

diff --git a/benchmarks/infraswe/README.md b/benchmarks/infraswe/README.md
@@ -0,0 +1,39 @@
+# InfraSWE Draft: common-shape BF16 GEMV
+
+This directory binds the common LLaMA/GPT-NeoX GEMV benchmark to AstrAI as an
+explicit repository target. AstrAI is not one of InfraSWE v0.5's built-in
+projects, so selecting a built-in default would score the change against the
+wrong host contract. The Draft remains `D3-contract-proposed`; it does not
+claim maintainer review, sealing, hidden-probe completion, or an official
+ProjectFit.
+
+The Draft was validated and resolved before opening the PR with InfraSWE commit
+`811bc775ed5b3a6ec853219245f3469f78818020`:
+
+```bash
+PYTHONPATH=src .venv/bin/infraswe draft validate \
+  /path/to/AstrAI/benchmarks/infraswe/astrai-gemv-common-draft.json
+
+PYTHONPATH=src .venv/bin/infraswe draft resolve \
+  --local-draft \
+  /path/to/AstrAI/benchmarks/infraswe/astrai-gemv-common-draft.json \
+  --output /tmp/astrai-gemv-common-draft-resolution.json
+```
+
+The candidate and contract digests bind the ordered source, tests,
+documentation, and raw benchmark evidence. The required comparison cell is one
+NVIDIA L20 (`sm_89`); optional A100 and H100 cells are explicitly untested.
+Compilation happens before timed cases. The checked-in kernel suite uses 21
+paired/interleaved samples, while the distinct-weight synthetic chains include
+the guarded Python dispatcher. The extended portfolio adds Qwen2-7B, LLaMA 3
+70B, and OPT-1.3B shapes at M=1/2/4/8, plus an interleaved
+128-thread-versus-256-thread CTA comparison. Neither suite is presented as
+whole-model throughput.
+
+Applying InfraSWE's frozen `project-fit-kernel-v0.5` formula to the visible
+evidence yields a diagnostic ProjectFit of **92.64/100** and BenchmarkTrust of
+**95.87/100**. The machine-readable rationale is
+`benchmarks/results/gemv_common_l20_sm89_infraswe_score.json`. Both numbers are
+non-official: official ProjectFit remains unresolved until the Draft is sealed,
+five or more fresh-process replays and system traces exist, hidden probes are
+complete, and the evidence manifest is verified.
diff --git a/benchmarks/infraswe/astrai-gemv-common-draft.json b/benchmarks/infraswe/astrai-gemv-common-draft.json
@@ -0,0 +1,103 @@
+{
+  "schema_version": "0.5",
+  "draft": {
+    "id": "astrai-common-shape-bf16-gemv-v1",
+    "revision": 2,
+    "state": "D3-contract-proposed",
+    "created_by": "0z5a"
+  },
+  "target": {
+    "mode": "repository",
+    "repository": "https://github.com/ViperEkura/AstrAI",
+    "revision": "sha256:59c45de7cb0356e9661cfb73d38bd11c888a8b8d564e9123a90c9fa3910cc498",
+    "project_profile_sha256": "sha256:953067c47f5298f06846c8fe873db55afae49b6588a81926e49d9ef6b0994e08"
+  },
+  "candidate": {
+    "kind": "git-diff",
+    "revision": "sha256:d9ae3993caf937517b121d4fe855f2c5172cf38f69aa4bda10bbb8daad52a730",
+    "intent": "add-fastpath",
+    "implementation_kind": "cuda-native",
+    "entrypoints": [
+      "astrai.extension.bf16_gemv",
+      "astrai.extension.backend.linear",
+      "scripts/tools/benchmark_gemv_common.py"
+    ],
+    "operator_family": "dense-gemm",
+    "phase": "inference",
+    "backend": "cuda",
+    "primary_host_candidate": "astrai"
+  },
+  "baseline": {
+    "mode": "target-head",
+    "revision": "sha256:59c45de7cb0356e9661cfb73d38bd11c888a8b8d564e9123a90c9fa3910cc498"
+  },
+  "deployment": {
+    "workload_portfolio": {
+      "id": "astrai-common-shape-gemv-chains-v1",
+      "sha256": "sha256:c3cf734200844d7b8308a4803f456cd9a3f0d7e19839e7d3e74778f6d1307e4d",
+      "path": "docs/benchmarks/gemv_common_extended_chain_l20_sm89.json"
+    },
+    "required_cells": [
+      "cuda-sm89-nvidia-l20"
+    ],
+    "optional_cells": [
+      "cuda-sm80-a100",
+      "cuda-sm90-h100"
+    ],
+    "request_or_step_protocol": {
+      "id": "astrai-gemv-common-benchmark-v1",
+      "sha256": "sha256:74808a77e551ffa5f77e3256930f2647ebd080828cfdbdcc2770f9d1fa298ff6",
+      "path": "docs/developer/decode_linear_benchmark.md"
+    }
+  },
+  "retrieval": {
+    "enabled": true,
+    "corpus_cutoff": "2026-09-02T09:45:00Z",
+    "sources": [
+      "target-code",
+      "merged-prs",
+      "rejected-prs",
+      "review-comments",
+      "ci-failures"
+    ],
+    "precedent_set_sha256": "sha256:9034271948fd791fb53d5099c2bc246acfd4e27e45926857220b48fd841389b9"
+  },
+  "acceptance_contract": {
+    "status": "proposed",
+    "path": "tests/extension/test_gemv.py,tests/extension/test_linear_dispatch.py",
+    "sha256": "sha256:15ffa9e1378d7cb1b06f018f9e1aac16bf56f14446271d51129f86c13f3827bb",
+    "probe_set_sha256": "sha256:beafc95a2a4aaa587079f1938db8c0fb115d63c2b0c3b185debc9a6132397653",
+    "hidden_probe_policy_sha256": "sha256:e6bdf6b5a307c9ebdaf643d18ae88b51c6e554d7d9df4427c2f7ba0eb5cc8e93"
+  },
+  "project_objectives": {
+    "edge_ecosystem_policy": "experimental",
+    "profile_set_sha256": "sha256:22da954f941ba912f60d79b7a58121436d4ba97e64f0182f66b8274080c3a75a"
+  },
+  "benchmark_loop": {
+    "fast_stage_max_official_fraction": 0.05,
+    "affected_stage_max_official_fraction": 0.2,
+    "official_replays": 7,
+    "early_exit_on_hard_gate": true,
+    "affected_case_selection": "required",
+    "benchmark_budget_policy_id": "draft-staged-budget-v0.5",
+    "evidence_policy_id": "v0.4-evidence-ladder-plus-seal-v0.5",
+    "precompile": {
+      "mode": "auto",
+      "trigger": "when-compilation-required",
+      "cache_policy": "content-addressed-evidence-identity",
+      "cache_miss_action": "precompile-before-timed-cases",
+      "timing_phases": [
+        "precompile",
+        "cold-start",
+        "steady-state"
+      ],
+      "steady_state_compile_allowed": false
+    }
+  },
+  "scoring": {
+    "formula_template_id": "project-fit-kernel-v0.5",
+    "provisional_scoring_allowed": true,
+    "official_scoring_requires_seal": true,
+    "project_season": "astrai-2026q3"
+  }
+}
diff --git a/benchmarks/results/gemv_common_l20_sm89_infraswe_score.json b/benchmarks/results/gemv_common_l20_sm89_infraswe_score.json
@@ -0,0 +1,122 @@
+{
+  "schema_version": "0.5",
+  "score_kind": "diagnostic-project-fit",
+  "score_is_official": false,
+  "draft_id": "astrai-common-shape-bf16-gemv-v1",
+  "draft_state": "D3-contract-proposed",
+  "formula_template_id": "project-fit-kernel-v0.5",
+  "diagnostic_project_fit_100": 92.6415162827573,
+  "component_values": {
+    "evolutionary_maintainability": 0.8835232637338853,
+    "project_contract_fit": 1.0,
+    "performance_reuse_utilization": 0.9515978740065268,
+    "operational_fit": 0.8438741102892859
+  },
+  "component_floors": {
+    "evolutionary_maintainability": 0.6,
+    "project_contract_fit": 0.6,
+    "performance_reuse_utilization": 0.4,
+    "operational_fit": 0.6
+  },
+  "subcomponent_inputs": {
+    "evolutionary_maintainability": {
+      "evolution": 0.78,
+      "locality": 0.9,
+      "tests": 1.0,
+      "failure": 1.0,
+      "contract": 0.9
+    },
+    "project_contract_fit": {
+      "integration": 1.0,
+      "interface": 1.0,
+      "lifecycle": 1.0,
+      "buildtest": 1.0,
+      "policy": 1.0
+    },
+    "performance_reuse_utilization": {
+      "attainment": 1.0,
+      "coverage": 0.82,
+      "retention": 1.0,
+      "family": 1.0,
+      "compile": 1.0
+    },
+    "operational_fit": {
+      "replay": 0.9,
+      "load": 0.82,
+      "resource": 0.9,
+      "coldsteady": 0.75
+    }
+  },
+  "input_rationale": {
+    "evolution": "The candidate extends the merged AstrAI GEMV stack with a rejected broad half-CTA experiment and an evidence-gated exact selector, but has no upstream maintenance history yet.",
+    "locality": "The CUDA change remains inside the GEMV launcher and one cooperative kernel template; dispatcher, benchmark, tests, and documentation are separately scoped.",
+    "tests": "The complete CPU-capable AstrAI suite passed (625 passed, 185 skipped) and the built L20 extension passed 140 focused GPU tests.",
+    "failure": "Unaligned K, unsupported M, training/autograd, unmeasured shapes, other architectures, losing half-CTA cells, and sub-threshold automatic bands fail closed to the existing kernel or PyTorch.",
+    "contract": "The benchmark and tests are digest-bound, but the D3 acceptance contract has not received maintainer review or a seal.",
+    "integration": "The fast path remains behind AstrAI's existing bf16_gemv primitive and linear backend.",
+    "interface": "No public signature or environment-variable contract changes; ASTRAI_GEMV modes remain compatible.",
+    "lifecycle": "The path remains inference-only, uses the current CUDA stream, fuses optional bias, and is CUDA Graph safe.",
+    "buildtest": "The SM89 extension built all 128/256-thread row specializations without spills, Ruff format/import checks passed, and both local and L20 tests passed.",
+    "policy": "The candidate adds no dependency and enables automatic dispatch only for measured architecture/M/(N,K) cells that clear per-shape and chain gates.",
+    "attainment": "Every newly enabled Qwen2, LLaMA 3 70B, and OPT chain exceeds the 3% gate; losing M/family combinations are explicitly excluded.",
+    "coverage": "Twenty-two traditional shapes and seven model families were measured in the required NVIDIA L20/SM89 cell; optional A100/SM80 and H100/SM90 cells remain untested and fall back in auto mode.",
+    "retention": "The exact row/shape half-CTA selector preserves the 256-thread fallback and existing warp-tiled bands; direct A/B evidence and the regression suite passed.",
+    "family": "Twenty-two LLaMA, Qwen2, OPT, and GPT-NeoX matrix shapes plus seven dependent synthetic chains cover M=1/2/4/8, without claiming a whole-model benchmark.",
+    "compile": "AOT CUDA compilation completed before timed cases; no compilation is permitted in steady-state timing.",
+    "replay": "Paired order alternation, warmup, 21 samples, median and p90 are captured, but not seven fresh processes.",
+    "load": "Distinct-weight dependent chains cover seven model families and include Python dispatch; a real end-to-end service workload was not run.",
+    "resource": "GPU5 memory and utilization were checked before and after; the co-resident idle AstrAI service remained allocated and is disclosed.",
+    "coldsteady": "Precompile and steady-state phases are separated, but cold-start latency is not part of the checked-in result."
+  },
+  "benchmark_trust": {
+    "formula_version": "benchmark-trust-v0.5",
+    "status": "scored",
+    "score_100": 95.87315155141826,
+    "components": {
+      "reproducibility": 1.0,
+      "evidence": 1.0,
+      "statistics": 0.9,
+      "environment": 0.9
+    },
+    "failure_codes": [
+      "DRAFT_UNSEALED",
+      "FRESH_PROCESS_REPLAY_INCOMPLETE",
+      "OPTIONAL_CELLS_UNTESTED"
+    ]
+  },
+  "official_project_fit": {
+    "status": "unresolved",
+    "score_100": null,
+    "failure_codes": [
+      "DRAFT_SEAL_MISSING",
+      "FRESH_PROCESS_REPLAYS_BELOW_MINIMUM",
+      "SYSTEM_TRACE_EVIDENCE_MISSING",
+      "HIDDEN_PROBES_INCOMPLETE",
+      "EVIDENCE_MANIFEST_UNVERIFIED"
+    ]
+  },
+  "comparison_cell": {
+    "target_project_profile_sha256": "sha256:953067c47f5298f06846c8fe873db55afae49b6588a81926e49d9ef6b0994e08",
+    "target_repository_or_baseline_sha256": "sha256:59c45de7cb0356e9661cfb73d38bd11c888a8b8d564e9123a90c9fa3910cc498",
+    "change_intent": "add-fastpath",
+    "semantic_contract_sha256": "sha256:74808a77e551ffa5f77e3256930f2647ebd080828cfdbdcc2770f9d1fa298ff6",
+    "acceptance_contract_sha256": "sha256:15ffa9e1378d7cb1b06f018f9e1aac16bf56f14446271d51129f86c13f3827bb",
+    "probe_set_sha256": "sha256:beafc95a2a4aaa587079f1938db8c0fb115d63c2b0c3b185debc9a6132397653",
+    "workload_portfolio_sha256": "sha256:c3cf734200844d7b8308a4803f456cd9a3f0d7e19839e7d3e74778f6d1307e4d",
+    "performance_target_sha256": "sha256:22da954f941ba912f60d79b7a58121436d4ba97e64f0182f66b8274080c3a75a",
+    "required_deployment_cell_set_sha256": "sha256:6c00c84e931a5ad6fdaa5b4f7d4497c872530c75f8510431aff883006a4a2779",
+    "formula_template_id": "project-fit-kernel-v0.5",
+    "evidence_policy_id": "v0.4-evidence-ladder-plus-seal-v0.5",
+    "project_season": "astrai-2026q3",
+    "cross_project_ranking_allowed": false
+  },
+  "execution": {
+    "infraswe_commit": "811bc775ed5b3a6ec853219245f3469f78818020",
+    "draft_resolution_sha256": "952ee658e669126582b8eb9aeedede9f70dae6348ea0f1119d64a36d5da100fe",
+    "infraswe_draft_engine_tests": "53 passed",
+    "astrai_tests": "625 passed, 185 skipped",
+    "astrai_l20_gpu_tests": "140 passed",
+    "astrai_lint": "ruff format and import-order checks passed",
+    "sm89_build": "bf16_gemv target built successfully; 128/256-thread cooperative specializations used 26-52 registers, one barrier, and zero spills; M=4 warp tiling used 42 registers, zero barriers, and zero spills"
+  }
+}