feat(dcu-kernel-auto-opt): harness planner, gate policy and DKAO tooling

metainfer/tasks/dcu_kernel_auto_opt/WORKFLOW.md

@@ -50,7 +50,7 @@ metainfer/tasks/dcu_kernel_auto_opt/
 ### 2.1 新建任务（表单字段，见 form.yaml）
 
 关键字段：`operator`（Quantized GEMM）、`kernel_language`（HIP C++）、`target_hardware`（K500SM_AI/gfx928）、
-`dtype`（INT8 W8A8）、`agent_framework`（ccb / dsh）、`agent_model`（ccb: Opus/Sonnet；dsh: deepseek-v4-flash）、`execution_mode`（Mock / Real INT8 W8A8 GEMM /
+`dtype`（INT8 W8A8）、`agent_framework`（ccb / dsh）、`agent_model`（ccb: Opus/Sonnet；dsh: deepseek-flash-4.1 / deepseek-v4-flash）、`execution_mode`（Mock / Real INT8 W8A8 GEMM /
 Generate & optimize / Infra smoke）、`target_repo_path`、`shape_assignment_mode`（AI automatic / Manual by GPU）、
 `shape_scope`（All API shapes / Selected shapes only）、`shape_config`、`max_iterations`、`minimum_improvement_percent`、`extra_notes`。
 
@@ -177,11 +177,25 @@ TP8 另有 wq_b/wo_b (1024,4096)、gate_up (4096,512)、down (256,4096)、indexe
   注意 TP1 目录前端可选、契约仍只收 tp_size 4/8，选 TP1 会同样在 prepare 被拒。
 - **模型目录 TP8 大 prefill 边界（2026-08-27）**：Hy3/MiniMax M3/GLM5.2 的 TP8 形状支持
   M=4096 优化（"Selected shapes only" 手动选择，如 minimax 任务）。契约加
-  `MODEL_TP8_EXTRA_OPTIMIZATION_M_VALUES=(4096,)`（**不进** `DEFAULT_OPTIMIZATION_SHAPES`，
+  `MODEL_TP8_EXTRA_OPTIMIZATION_M_VALUES`（**不进** `DEFAULT_OPTIMIZATION_SHAPES`，
   默认 42 不变，串行验证 fallback 不受影响）；前端三个模型的 TP8 topology 加
-  `mValues: [2,16,3072,4096]`；基线表 15 条 `(8,4096,…)` 已实测
+  `mValues`（用 `W8A8_MODEL_TP8_M_VALUES` 常量）；基线表 15 条 `(8,4096,…)` 已实测
   （`tools/baseline/bench_triton_tp8_m4096.py`，int8_utils.matmul_kernel + CUDA-graph replay，
   结果存 `tools/baseline/tp8_m4096_graph.json`）。DeepSeek TP8 不加 M=4096。
+- **模型目录 TP8 短 decode 边界（2026-09-20）**：Hy3/MiniMax M3/GLM5.2 的 TP8 形状再加
+  M=8，即 `MODEL_TP8_EXTRA_OPTIMIZATION_M_VALUES=(8, 4096)`；前端
+  `W8A8_MODEL_TP8_M_VALUES = [2, 8, 16, 3072, 4096]`（DeepSeek TP1/TP8 仍是
+  `W8A8_M_VALUES=[2,16,3072]`，TP4 仍是 `W8A8_TP4_M_VALUES=[2,16,3072,4096]`）。
+  基线表 15 条 `(8,8,…)` 在 worker29/gfx928 实测三次
+  （`tools/baseline/bench_triton_tp8_m8.py --m 8 --mode graph`，协议同 2026-08-12 目录：
+  CUDA-graph replay、hot cache、warmups=10/samples=20/launches=5、BM16/BN32/BK256），
+  公布值取「每次 pass median 再取 median」；原始数据
+  `tools/baseline/tp8_m8_graph_run{1,2,3}.json`，汇总 `tp8_m8_graph_merged.json`。
+  12/15 形状三次 pass 极差 ≤1.6%（共享卡上 pass 2 的 minimax shared_down_proj 被同租户
+  干扰到 33.76 µs，取 median 后为 19.98 µs）。DeepSeek TP8 不加 M=8。
+- **面板/契约/基线三层一致性测试**：`tests/test_shape_catalog_panel.py` 用 node 真实求值
+  `static/dkao-shape-input.js` 的 `MODEL_WORKLOADS`，断言面板给出的每个 TP4/TP8 shape
+  都能过契约校验且能查到基线（否则任务要到 baseline 阶段才报 ValueError）。
 - `MIN_M=1, MAX_M=4096`；`WORKSPACE_BUDGET_BYTES=16MB`。
 - **M=4096 时大部分 (N,K) 的 split-K workspace 容量为 0** → 大 M kernel 必须走 2D M-tile 路径，
   不能依赖 split-K workspace。
@@ -257,7 +271,8 @@ TP8 另有 wq_b/wo_b (1024,4096)、gate_up (4096,512)、down (256,4096)、indexe
    **规范 skill 以 `metainfer/tasks/dcu_kernel_auto_opt/skills/` 内的副本为种子**（int8-w8a8 家族，
    2026-08-27 起随插件一起维护）：新机器上 `sync_skill_libraries()`（或 WebUI 同步按钮）会自动
    把缺失的 skill 补种进 `~/.dsh/skills/`，再镜像到 `~/.claude/skills/`；已有的库 skill 不被覆盖。
-   baseline 测量工具在 `tools/baseline/`（`bench_triton_tp4/tp8_m4096.py` + `int8_utils.py`）。
+   baseline 测量工具在 `tools/baseline/`（`bench_triton_tp4_m4096.py`、
+   `bench_triton_tp8_m4096.py`、`bench_triton_tp8_m8.py` + `int8_utils.py`）。
    kernel-repos 默认在 MetaInfer 同级（`METAINFER_KERNEL_REPOS` 可改），不随插件目录走。
 5. 改动任何行为后：更新本文件相关段落 + 跑 tests + 用真实任务验证（优先在 zth_meta 里）。
 

metainfer/tasks/dcu_kernel_auto_opt/api/int8w8a8gemm/int8_w8a8_gemm_api.py

@@ -193,14 +193,19 @@ def _merge_kn_tables(
 # the original three M values, so each TP4 operator gets exactly one M=4096
 # shape while TP8 defaults are unchanged.
 TP4_EXTRA_OPTIMIZATION_M_VALUES: Final[tuple[int, ...]] = (4096,)
-# Model-catalog TP8 large-prefill boundary added on 2026-08-27. The DeepSeek
-# TP8 default workload keeps the original three M values; Hy3 / MiniMax M3 /
-# GLM5.2 TP8 operators are additionally optimizable at M=4096 via
-# "Selected shapes only" (their (K,N) pairs are already in the allowed sets
-# and their Triton CUDA-graph baselines are measured). This constant is not
-# applied by _default_optimization_shapes(), so DEFAULT_OPTIMIZATION_SHAPES
+# Model-catalog TP8 extra M values. The DeepSeek TP8 default workload keeps
+# the original three M values; Hy3 / MiniMax M3 / GLM5.2 TP8 operators are
+# additionally optimizable at these M values via "Selected shapes only"
+# (their (K,N) pairs are already in the allowed sets and their Triton
+# CUDA-graph baselines are measured). This constant is not applied by
+# _default_optimization_shapes(), so DEFAULT_OPTIMIZATION_SHAPES
 # (DeepSeek-only) and its serial-validation fallback scope are unchanged.
-MODEL_TP8_EXTRA_OPTIMIZATION_M_VALUES: Final[tuple[int, ...]] = (4096,)
+#   M=4096: model-catalog large-prefill boundary added on 2026-08-27.
+#   M=8:    model-catalog short decode boundary added on 2026-09-20; the
+#           Triton Graph baselines for the 15 catalog (K, N) pairs were
+#           measured on worker29/gfx928 with tools/baseline/
+#           bench_triton_tp8_m8.py (raw passes in tp8_m8_graph_run{1,2,3}.json).
+MODEL_TP8_EXTRA_OPTIMIZATION_M_VALUES: Final[tuple[int, ...]] = (8, 4096)
 
 
 def _default_optimization_shapes() -> tuple[dict[str, int | str], ...]:

metainfer/tasks/dcu_kernel_auto_opt/assets/w8a8_bench.py

@@ -8,8 +8,11 @@
 import json
 import math
 import os
+import re
 import statistics
+import subprocess
 import sys
+import time
 from pathlib import Path
 
 try:
@@ -18,6 +21,159 @@
     torch = None  # type: ignore[assignment]
 
 
+#: Device admission rule for anything that measures: VRAM <= 90% and HCU == 0.
+#: This harness is the only thing that runs the candidate kernel, and an agent
+#: can invoke it directly from its own shell, so the rule is enforced here too —
+#: not only in the orchestrator that happens to launch us. A number taken while
+#: another workload shares the card is noise (the same unchanged shape has been
+#: seen at 884us -> 3047us), so we refuse to produce one.
+GATE_EXIT_CODE = 75
+_GATE_DRM_ROOT = Path("/sys/class/drm")
+_GATE_SMI_ROW = re.compile(
+    r"^(?P<index>\d+)\s+[\d.]+C\s+[\d.]+W\s+\S+\s+[\d.]+W\s+"
+    r"(?P<vram>[\d.]+|N/A|n/a)%\s+(?P<util>[\d.]+|N/A|n/a)%"
+)
+
+
+def _gate_number(value):
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        return None
+
+
+def _gate_hy_smi() -> dict:
+    """``{gpu: {vram_percent, util_percent}}`` from hy-smi; empty if unreadable."""
+    for binary in ("hy-smi", "rocm-smi"):
+        try:
+            proc = subprocess.run([binary], capture_output=True, text=True,
+                                  timeout=10)
+        except (OSError, subprocess.SubprocessError):
+            continue
+        if proc.returncode != 0 or not proc.stdout.strip():
+            continue
+        out = {}
+        for line in proc.stdout.splitlines():
+            match = _GATE_SMI_ROW.match(line.strip())
+            if not match:
+                continue
+            out[int(match.group("index"))] = {
+                "vram_percent": _gate_number(match.group("vram").strip("%")),
+                "util_percent": _gate_number(match.group("util").strip("%")),
+            }
+        if out:
+            return out
+    return {}
+
+
+def _gate_sysfs() -> dict:
+    """Driver sysfs fallback: works when the management interface says N/A."""
+    out = {}
+    cards = []
+    try:
+        entries = sorted(_GATE_DRM_ROOT.iterdir(), key=lambda p: p.name)
+    except OSError:
+        return out
+    for entry in entries:
+        if not re.fullmatch(r"card\d+", entry.name):
+            continue
+        device = entry / "device"
+        try:
+            uevent = (device / "uevent").read_text(encoding="utf-8")
+        except OSError:
+            continue
+        if not any(driver in uevent for driver in ("hycu", "amdgpu", "hydcu")):
+            continue
+        if not (device / "mem_info_vram_total").is_file():
+            continue
+        cards.append(device)
+    for index, device in enumerate(cards):
+        row = {}
+        try:
+            row["util_percent"] = float(
+                (device / "gpu_busy_percent").read_text().strip())
+        except (OSError, ValueError):
+            pass
+        try:
+            used = float((device / "mem_info_vram_used").read_text().strip())
+            total = float((device / "mem_info_vram_total").read_text().strip())
+            if total > 0:
+                row["vram_percent"] = round(100.0 * used / total, 2)
+        except (OSError, ValueError, ZeroDivisionError):
+            pass
+        if row:
+            out[index] = row
+    return out
+
+
+def _gate_violation(device: int) -> str:
+    """``""`` when the device passes, else why it does not."""
+    from_smi = _gate_hy_smi().get(device) or {}
+    complete = (from_smi.get("vram_percent") is not None
+                and from_smi.get("util_percent") is not None)
+    readings = from_smi if complete else dict(_gate_sysfs().get(device) or {})
+    if not complete:
+        for key in ("vram_percent", "util_percent"):
+            if readings.get(key) is None and from_smi.get(key) is not None:
+                readings[key] = from_smi[key]
+    vram = readings.get("vram_percent")
+    util = readings.get("util_percent")
+    if vram is None or util is None:
+        return "device state unavailable (no HCU/VRAM reading)"
+    if vram > 90.0:
+        return f"VRAM {vram:.0f}% > 90%"
+    if util > 0.0:
+        return f"HCU {util:.0f}% > 0%"
+    return ""
+
+
+def enforce_measurement_gate(device: int | None = None) -> None:
+    """Wait for ``VRAM <= 90% and HCU == 0``; exit if it never happens.
+
+    ``METAINFER_GPU_PREFLIGHT=0`` is the deliberate opt-out. Waiting is
+    deliberately bounded (``METAINFER_GATE_WAIT_SECONDS`` x
+    ``METAINFER_GATE_MAX_WAITS``, default 30 min x 48 = 24 h): a shared card
+    must never silently turn into a measurement, and waiting forever is not an
+    answer either.
+    """
+    if str(os.environ.get("METAINFER_GPU_PREFLIGHT", "")).strip().lower() in {
+            "0", "false", "no", "off"}:
+        return
+    if device is None:
+        raw = str(os.environ.get("HIP_VISIBLE_DEVICES")
+                  or os.environ.get("ROCR_VISIBLE_DEVICES") or "0")
+        device = int(raw.split(",")[0].strip() or 0)
+    try:
+        wait_s = int(os.environ.get("METAINFER_GATE_WAIT_SECONDS", "") or 1800)
+    except ValueError:
+        wait_s = 1800
+    try:
+        max_waits = int(os.environ.get("METAINFER_GATE_MAX_WAITS", "") or 48)
+    except ValueError:
+        max_waits = 48
+    for attempt in range(1, max_waits + 1):
+        reason = _gate_violation(int(device))
+        if not reason:
+            return
+        if attempt > max_waits:
+            break
+        print(json.dumps({
+            "measurement_gate": "blocked", "device": int(device),
+            "attempt": attempt, "max_waits": max_waits,
+            "wait_seconds": wait_s, "reason": reason,
+        }), file=sys.stderr)
+        if attempt == max_waits:
+            break
+        time.sleep(max(0.0, float(wait_s)))
+    reason = _gate_violation(int(device)) or "device never became idle"
+    print(json.dumps({
+        "measurement_gate": "give_up", "device": int(device),
+        "checks": max_waits, "reason": reason,
+    }), file=sys.stderr)
+    raise SystemExit(GATE_EXIT_CODE)
+
+
+
 def load_module(module_path: Path, name: str):
     spec = importlib.util.spec_from_file_location(
         name, module_path
@@ -331,6 +487,11 @@ def _env_default(name: str, fallback: int) -> int:
         }, sort_keys=True))
         return 0
 
+    # Everything below initialises the device and measures on it. CPU-only runs
+    # (--self-test, --prepare-reference) returned above, so they are never
+    # blocked; every device-touching run has to pass the gate first.
+    enforce_measurement_gate()
+
     source = args.source.resolve()
     fixed_contract = source / "int8_w8a8_gemm_api.py"
     if fixed_contract.is_file():

metainfer/tasks/dcu_kernel_auto_opt/bridge/dsh/README.md

@@ -32,8 +32,9 @@ prompt 从 stdin 传入，stdout 输出逐行 stream-json 事件
 WebUI new-task 表单中 dcu-kernel-auto-opt 增加 **Agent framework** 选择：
 
 - `ccb` → Claude Code（模型 `Sonnet` / `Opus`）
-- `dsh` → DeepSeek Harness（模型 `deepseek-v4-flash`，即
-  `deepseek/deepseek-v4-flash-0731`）
+- `dsh` → DeepSeek Harness：
+  - `deepseek-flash-4.1` → `deepseek/deepseek-flash`（**默认**，4.1 Flash）
+  - `deepseek-v4-flash` → `deepseek/deepseek-v4-flash-0731`（旧 pinned 构建，可复现历史运行）
 
 orchestrator CLI 在 `agent_framework=dsh` 时自动把 `claude_bin` 指向本目录的
 `dsh_agent.py`（`--claude-bin` 显式传入时优先），无需改动 MetaInfer 共享代码。
@@ -55,15 +56,16 @@ runtime 只内置 `deepseek-official` adapter（`llm-deepseek`），它通过
 
 - provider：`deepseek-official`（runtime adapter；TokenHub 经 base_url 覆盖）
 - baseURL：`https://tokenhub.tencentmaas.com/plan/v3`
-- 模型：`deepseek/deepseek-v4-flash-0731`（默认）
+- 模型：`deepseek/deepseek-flash`（默认，= 表单里的 `deepseek-flash-4.1`）；
+  旧 pinned 构建 `deepseek/deepseek-v4-flash-0731`（= `deepseek-v4-flash`）
 - API key：`~/.dsh/.credentials.yaml` 的 `TENCENT_API_KEY`
 
 `dsh_agent.py` 的环境变量：
 
 | 变量 | 默认 | 说明 |
 |---|---|---|
 | `DSH_AGENT_PROVIDER` | `deepseek-official` | DSH runtime provider（dev runtime 仅内置此 adapter） |
-| `DSH_AGENT_MODEL` | `deepseek/deepseek-v4-flash-0731` | 模型覆盖 |
+| `DSH_AGENT_MODEL` | `deepseek/deepseek-flash` | 默认模型覆盖（只改 `deepseek-flash-4.1` 这个默认标签的 id） |
 | `DSH_AGENT_BASE_URL` | `https://tokenhub.tencentmaas.com/plan/v3` | 模型端点（`DEEPSEEK_BASE_URL` 优先） |
 | `TENCENT_API_KEY` / `DEEPSEEK_API_KEY` | 凭据文件兜底 | API key |
 | `DSH_AGENT_CORDIS` | 本目录 `cordis.yml` | 自定义组合路径 |

metainfer/tasks/dcu_kernel_auto_opt/bridge/dsh/dsh_agent.py

@@ -27,7 +27,8 @@
     DSH_AGENT_PROVIDER      provider name for the DSH runtime
                             (default: deepseek-official — the only adapter
                             shipped by the dev-checkout runtime)
-    DSH_AGENT_MODEL         model override (default deepseek/deepseek-v4-flash-0731)
+    DSH_AGENT_MODEL         model override (default deepseek/deepseek-flash,
+                            the 4.1 Flash label deepseek-flash-4.1)
     DSH_AGENT_BASE_URL      model endpoint; falls back to DEEPSEEK_BASE_URL
                             (default: https://tokenhub.tencentmaas.com/plan/v3)
     TENCENT_API_KEY         preferred API key (matches ~/.dsh/settings.yaml)
@@ -71,9 +72,11 @@ def default_provider() -> str:
 
 
 def default_model() -> str:
+    """Default DSH model id (the 4.1 Flash build) unless DSH_AGENT_MODEL pins
+    another one. Matches orchestrator/config.py DSH_DEFAULT_MODEL_ID."""
     return (
-        os.environ.get("DSH_AGENT_MODEL", "deepseek/deepseek-v4-flash-0731")
-        .strip() or "deepseek/deepseek-v4-flash-0731"
+        os.environ.get("DSH_AGENT_MODEL", "deepseek/deepseek-flash")
+        .strip() or "deepseek/deepseek-flash"
     )
 
 
@@ -144,13 +147,20 @@ def default_api_key() -> str:
 
 
 # --------------------------------------------------------------------------- #
-# Model mapping (legacy Claude labels -> this host's DSH model ids)
+# Model mapping (legacy Claude labels / form labels -> this host's DSH ids)
 # --------------------------------------------------------------------------- #
 
+#: Form labels produced by both MetaInfer task frontends (dcu-kernel-auto-opt's
+#: agent_model field and harness-evolve's evolve_model field). Kept in sync with
+#: orchestrator/config.py (DSH_MODEL_IDS) — the orchestrator resolves the label
+#: itself, so this map only serves callers that hand the wrapper a bare label.
 _MODEL_MAP = {
-    "opus": "deepseek-v4-pro",
-    "sonnet": "deepseek-v4-flash",
-    "haiku": "deepseek-v4-flash",
+    "deepseek-flash-4.1": "deepseek/deepseek-flash",
+    "deepseek-flash": "deepseek/deepseek-flash",
+    "deepseek-v4-flash": "deepseek/deepseek-v4-flash-0731",
+    "opus": "deepseek/deepseek-v4-flash-0731",
+    "sonnet": "deepseek/deepseek-v4-flash-0731",
+    "haiku": "deepseek/deepseek-v4-flash-0731",
 }
 
 
@@ -160,11 +170,8 @@ def map_model(requested: Optional[str]) -> str:
     key = requested.strip().lower()
     if key in _MODEL_MAP:
         return _MODEL_MAP[key]
-    if key == "deepseek-v4-flash":
-        # Bare user-facing label -> the pinned host model id.
-        return default_model()
     if key.startswith("deepseek"):
-        # Full model id (e.g. deepseek/deepseek-v4-flash-0731): pass through.
+        # Full model id (e.g. deepseek/deepseek-flash): pass through.
         return key
     # Unknown label: keep the caller's intent but stay on a known model id.
     return default_model()
@@ -294,6 +301,13 @@ def _headless_backend_main(args: Any, prompt: str) -> int:
     sub-agent is still working so SubAgentManager's no-output watchdog never
     mistakes a long task for a stuck process; the final assistant text is
     emitted as the last message, matching the ccb stream-json contract.
+
+    Model selection caveat: the ``dsh --profile headless`` app takes no
+    ``--model`` flag, so it runs whatever the user's DSH settings document
+    (``~/.dsh/settings.yaml`` -> ``agent-default-model``) selects — the host
+    default, not ``--model``. Emit the requested model in the system event so
+    the mismatch is visible instead of silent. The SDK backend below does honour
+    ``--model``.
     """
     import subprocess
     import tempfile

metainfer/tasks/dcu_kernel_auto_opt/bridge/dsh/run_dsh_task.sh

@@ -12,7 +12,8 @@
 #
 # Optional env:
 #   DSH_AGENT_MAX_TOKENS   per-request output cap (default 65536)
-#   DSH_AGENT_MODEL        model override (default deepseek/deepseek-v4-flash-0731)
+#   DSH_AGENT_MODEL        model override (default deepseek/deepseek-flash,
+#                          the 4.1 Flash build)
 #   METAINFER_GPU_IDS      restrict GPUs, e.g. 4,5,6,7
 
 set -euo pipefail

metainfer/tasks/dcu_kernel_auto_opt/bridge/dsh/tests/smoke_sdk.py

@@ -30,7 +30,7 @@ def main() -> int:
     WORK.mkdir(parents=True, exist_ok=True)
     cfg = DeepSeekHarnessConfig(
         provider="deepseek-official",
-        model="deepseek-v4-flash",
+        model="deepseek/deepseek-flash",
         max_tokens=2048,
         cwd=str(WORK),
         session_root=str(SESSION_ROOT),