Both versions receive original ONNX initializer arrays without reshaping. The current parser's arrays must match dtype/shape/bytes, and its extra Constant globals are added. The baseline parser has no initializers property; supplying this existing interpreter argument is a harness adapter, not a baseline implementation change.
parse: fresh ONNXParser.parse per sample; run: IRInterpreter.run including verification, planning, binding copies, NumPy computation and return copy. ORT reference generation, common initializer loading, external verification and comparisons are excluded. Each round starts with one untimed correctness qualification, then the configured warmups and timed repetitions. Sequential ABBA order, two rounds per version; warmup is per round. Host CPU, not QEMU.
-
执行计时覆盖完整 IRInterpreter.run，包括 IR 校验与输入/权重复制；不含 ONNX 解析和 ONNX Runtime 参考计算。ONNX → IR 解析单独计时。
"]
+ failures = []
+ if report.get("error"):
+ failures.append(f"阶段 {report.get('stage')}: {report['error']}")
+ for case in cases:
+ if case.get("passed") is not True:
+ failures.append(f"{case.get('name')}: FAIL；有检查失败或未完成")
+ for row in case.get("qemu", []):
+ if row.get("passed") is not True:
+ reason = row.get("error") or row.get("reason") or "尚未完成数值核验"
+ first = row.get("checkpoints", {}).get("first_divergence")
+ failures.append(f"{case.get('name')} / {row.get('graph')}/{row.get('optimization')}: "
+ f"{row.get('status')}; {reason}" + (f"；首次超限检查点 {first}" if first else ""))
+ for row in case.get("optimized_ir", []):
+ if row.get("passed") is not True:
+ failures.append(f"{case.get('name')} IR {row.get('graph')}/{row.get('optimization')}: "
+ f"{row.get('reason') or 'FAIL'}")
+ for row in report.get("invariants", []):
+ if row.get("passed") is not True:
+ failures.append(f"{row.get('name')} / {row.get('graph')}/{row.get('optimization')}: 隔离检查 FAIL")
+ if not report.get("passed") and not failures:
+ failures.append("尚未完成全部验收，请检查 report.json。")
+ if failures:
+ markdown += ["", "**失败／未完成项**", "", *[f"- {escape(item)}" for item in failures]]
+ html.append("
失败／未完成项
"
+ + "".join(f"
{escape(item)}
" for item in failures) + "
")
+
+ def append_detail(title, headers, rows, note):
+ md, body = _table_views(headers, rows)
+ md, body = _details_views(title, md + "\n\n" + note, body + f"
{escape(note)}
")
+ markdown.extend(["", md])
+ html.append(body)
+
+ graph_note = ("图指模型计算步骤及其连接关系：normal 只输出最终 logits；diagnostic 增加中间检查点以定位误差，含额外打包与传输开销。"
+ "none/basic/all 指 ScratchV IR 优化：none 不优化；basic 提前计算常量并删除不影响输出的计算；"
+ "all 再加入局部指令简化、乘加融合和循环内重复计算外提。它们不是 C 编译优化等级，本探测的 C 编译固定使用 -O2。")
+ time_note = (graph_note + " QEMU 墙钟时间包含启动、目标程序计算、输出打包/串口传输和退出；超时还含清理，"
+ "不含主机输入准备或输出解码。未测得的时长不按零计；平均/最大值仅统计有测量值的样本。"
+ "七种输入不是重复采样，不能据此推导稳定加速比，也不设置性能通过阈值。"
+ f"成功交叉编译记录累计 {format_seconds(timing['cross_compile_seconds_total'])} 秒"
+ f"（{timing['cross_compile_completed_count']}/{timing['cross_compile_planned_count']} 次），"
+ "不含 ScratchV 解析、IR 优化、代码生成及失败构建；流程时间不含报告渲染。")
+ groups = []
for group in timing["groups"]:
successful, failed = group["successful"], group["failed"]
- failed_count = group["attempted_count"] - group["status_counts"]["success"]
- cells = [f"{group['graph']}/{group['optimization']}",
- f"{group['attempted_count']}/{group['planned_count']}",
- f"{group['status_counts']['success']}/{failed_count}", str(group['not_attempted_count']),
- f"{format_seconds(successful['mean_seconds'])}/{format_seconds(successful['max_seconds'])}",
- f"{format_seconds(failed['mean_seconds'])}/{format_seconds(failed['max_seconds'])}",
- format_seconds(group['all_measured']['total_seconds'])]
- group_rows.append(cells)
- lines.append("| " + " | ".join(cells) + " |")
- lines += ["", "平均/最大值只统计有实测时长的样本；失败数包含尚未启动但已尝试的条目。"
- "七种输入不是重复采样，不据此宣称稳定加速比。", "",
- "FP32, L=256, random weights, 29 checkpoints, strict max absolute error < 1e-5.",
- "The C cross-compiler is Zig/LLVM; this does not validate the legacy scalar assembly selector.",
- "IR optimization levels: none/basic/all. QEMU: none/all, normal and diagnostic graphs.", ""]
- if report.get("error"):
- lines += [f"Error: {report['error']}", ""]
- lines += ["| Case | Graph | IR optimization | QEMU进程(s) | 超时设定(s) | Max abs | First divergent checkpoint | Status |",
- "|---|---|---|---:|---:|---:|---|---|"]
- execution_rows = []
- for case in report.get("cases", []):
+ groups.append([f"{group['graph']}/{group['optimization']}",
+ f"{group['attempted_count']}/{group['planned_count']}",
+ f"{group['status_counts']['success']}/{group['attempted_count'] - group['status_counts']['success']}",
+ group['not_attempted_count'],
+ f"{format_seconds(successful['mean_seconds'])}/{format_seconds(successful['max_seconds'])}",
+ f"{format_seconds(failed['mean_seconds'])}/{format_seconds(failed['max_seconds'])}",
+ format_seconds(group['all_measured']['total_seconds'])])
+ append_detail("图与优化级别：仿真耗时明细",
+ ["图/优化级别", "已尝试/计划", "通过/失败", "未尝试", "通过平均/最大(s)", "失败平均/最大(s)", "已测合计(s)"],
+ groups, time_note)
+ executions = []
+ for case in cases:
for row in case.get("qemu", []):
- cells = [case['name'], row['graph'], row['optimization'],
- format_seconds(row.get('qemu_process_wall_seconds')),
- format_seconds(row.get('timeout_seconds')), str(row.get('max_abs')),
- row.get('checkpoints', {}).get('first_divergence') or 'none', row['status']]
- execution_rows.append(cells)
- lines.append("| " + " | ".join(cells) + " |")
- markdown = "\n".join(lines) + "\n"
- details = []
- for case in report.get("cases", []):
- details.append(f"{escape(case['name'])}: "
- f"{'PASS' if case['passed'] else 'FAIL'}
"]
+ failures = []
if report.get("error"):
- lines += [f"Failed stage: `{report['stage']}`", "", "```text", report["error"], "```", ""]
- lines += ["| Case | Valid tokens | PyTorch / ORT max error | IR / ORT max error | First divergence | Result |",
- "|---|---:|---:|---:|---|---|"]
- for case in report.get("cases", []):
- pairs = [case.get(name, {}) for name in ("pytorch_vs_ort", "ir_vs_ort")]
- maxima = [max((row["max_abs"] for row in pair.get("checkpoints", [])
- if row.get("max_abs") is not None), default=None) for pair in pairs]
- values = [f"{value:.3e}" if value is not None else "n/a" for value in maxima]
- first = next((pair.get("first_divergence") for pair in pairs if pair.get("first_divergence")), "none")
- lines.append(f"| {case['name']} | {case['valid_length']} | {values[0]} | {values[1]} | "
- f"{first} | {'PASS' if case['passed'] else 'FAIL'} |")
- for case in report.get("cases", []):
- lines += ["", f"{case['name']}: checkpoint comparisons", "",
- "| Comparison | Checkpoint | Max absolute error | Result |", "|---|---|---:|---|"]
+ failures.append(f"阶段 {report.get('stage', 'unknown')}: {report['error']}")
+ for case in cases:
+ if case.get("passed") is not True:
+ first = next((case.get(pair, {}).get("first_divergence") for pair in
+ ("pytorch_vs_ort", "ir_vs_ort") if case.get(pair, {}).get("first_divergence")), None)
+ failures.append(f"{case.get('name', 'unknown')}: FAIL；"
+ + (f"首次超限检查点 {first}" if first else "有检查失败或尚未完成，请展开明细"))
+ for check in report.get("invariants", []):
+ if check.get("passed") is not True:
+ failures.append(f"{check.get('name')} / {check.get('backend')}: 隔离检查 FAIL")
+ if not report.get("passed") and not failures:
+ failures.append(f"阶段 {report.get('stage', 'unknown')}: 尚未完成全部验收")
+ if failures:
+ markdown += ["", "**失败／未完成项**", "", *[f"- {escape(item)}" for item in failures]]
+ html.append("
失败／未完成项
"
+ + "".join(f"
{escape(item)}
" for item in failures) + "
")
+
+ def append_detail(title, headers, rows, note="", *, opened=False):
+ md, body = _table_views(headers, rows)
+ if note:
+ md += "\n\n" + note
+ body += f"
{escape(note)}
"
+ md, body = _details_views(title, md, body, opened=opened)
+ markdown.extend(["", md])
+ html.append(body)
+
+ case_rows = []
+ for case in cases:
+ ordinary = case.get("ordinary_logits", {})
+ case_rows.append([case.get("name"), case.get("valid_length"),
+ ordinary.get("pytorch_vs_ort", {}).get("max_abs", "未统计"),
+ ordinary.get("ir_vs_ort", {}).get("max_abs", "未统计"),
+ "PASS" if case.get("passed") else "FAIL"])
+ append_detail("输入场景明细", ["场景", "有效 token", "ORT / PyTorch MaxAbs", "IR / ORT MaxAbs", "全部检查"], case_rows)
+ checkpoint_rows = []
+ for case in cases:
for pair in ("pytorch_vs_ort", "ir_vs_ort"):
for row in case.get(pair, {}).get("checkpoints", []):
- lines.append(f"| {pair} | {row['name']} | {row.get('max_abs')} | "
- f"{'PASS' if row['passed'] else 'FAIL'} |")
- lines += ["", ""]
- lines += ["", "| Invariant | Backend | Max absolute error | Result |", "|---|---|---:|---|"]
- for check in report.get("invariants", []):
- lines.append(f"| {check['name']} | {check['backend']} | {check.get('max_abs')} | "
- f"{'PASS' if check['passed'] else 'FAIL'} |")
- return "\n".join(lines) + "\n"
+ checkpoint_rows.append([case.get("name"), pair, row.get("name"), row.get("max_abs"),
+ "PASS" if row.get("passed") else row.get("reason") or "FAIL"])
+ for name, row in case.get("ordinary_logits", {}).items():
+ if "diagnostic_vs_ordinary" in name:
+ checkpoint_rows.append([case.get("name"), name, "logits", row.get("max_abs"),
+ "PASS" if row.get("passed") else row.get("reason") or "FAIL"])
+ append_detail("检查点与普通/诊断图一致性", ["场景", "比较", "检查点", "MaxAbs", "结果"], checkpoint_rows,
+ "normal 为只输出最终 logits 的普通图；diagnostic 额外输出中间检查点以定位误差。两种图都参与验收。")
+ append_detail("因果与补齐隔离检查", ["检查", "执行路径", "MaxAbs", "结果"],
+ [[row.get("name"), row.get("backend"), row.get("max_abs"),
+ "PASS" if row.get("passed") else "FAIL"] for row in report.get("invariants", [])])
+ footer = "完整诊断、位置明细、命令、环境与源码指纹保存在 report.json；原始数组随 CI artifact 提供。"
+ markdown.extend(["", footer])
+ html.append(f"
Official Qwen3 with random weights. All positions, including padding queries, are checked. "
- "Ordinary logits and diagnostic graphs must both pass. This is not a RISC-V or pretrained quality test.
"]
- if report.get("error"):
- sections.append(f"
{cell(report['stage'])}: {cell(report['error'])}
")
- for case in report.get("cases", []):
- case_status = "PASS" if case["passed"] else "FAIL"
- details = []
- for pair in ("pytorch_vs_ort", "ir_vs_ort"):
- comparison = case.get(pair, {})
- details.append(f"
{cell(pair)}
First divergence: "
- f"{cell(comparison.get('first_divergence') or 'none')}