Spaces:
Sleeping
Sleeping
File size: 15,812 Bytes
490d677 ceab523 490d677 65a157d 490d677 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 | from __future__ import annotations
import json
import sys
import tempfile
import zipfile
from dataclasses import dataclass
from pathlib import Path
from typing import Any, Dict, List, Mapping, Optional, Sequence, Tuple
def _bootstrap_src_path() -> None:
here = Path(__file__).resolve()
candidates = [parent / "src" for parent in (here.parent, *tuple(here.parents))]
for candidate in candidates:
if candidate.exists() and str(candidate) not in sys.path:
sys.path.insert(0, str(candidate))
return
_bootstrap_src_path()
from nsys_llm_explainer.queries import TraceDB # type: ignore
from nsys_llm_explainer.report import AnalysisOutputs, analyze, render_markdown, write_artifacts # type: ignore
@dataclass(frozen=True)
class SpaceBundle:
source_path: Path
source_kind: str
report: Dict[str, Any]
markdown: str
artifacts_dir: Path
artifact_paths: List[Path]
summary_rows: List[Dict[str, str]]
manifest_rows: List[Dict[str, str]]
findings_markdown: str
status_markdown: str
def _coerce_float(value: Any, default: float = 0.0) -> float:
try:
return float(value)
except Exception:
return float(default)
def _safe_text(value: Any, default: str = "-") -> str:
text = str(value).strip() if value is not None else ""
return text if text else default
def _safe_trace_name(report: Mapping[str, Any]) -> str:
trace_path = ((report.get("trace") or {}).get("path") or report.get("_source_name") or "")
return Path(str(trace_path)).name if trace_path else "unknown"
def _top_kernel_row(report: Mapping[str, Any]) -> Optional[Mapping[str, Any]]:
rows = ((report.get("metrics") or {}).get("top_kernels") or {}).get("kernels") or []
return rows[0] if rows else None
def _top_nccl_row(report: Mapping[str, Any]) -> Optional[Mapping[str, Any]]:
rows = ((report.get("metrics") or {}).get("nccl") or {}).get("ops") or []
return rows[0] if rows else None
def _format_ms(value: Any) -> str:
return "{:.3f} ms".format(_coerce_float(value))
def _format_us(value: Any) -> str:
return "{:.2f} us".format(_coerce_float(value))
def _format_pct(value: Any) -> str:
return "{:.1f}%".format(_coerce_float(value))
def _bottleneck_sentence(report: Mapping[str, Any]) -> str:
metrics = report.get("metrics") or {}
total_gpu_ms = _coerce_float((metrics.get("top_kernels") or {}).get("total_kernel_time_ns")) / 1_000_000.0
top_kernel = _top_kernel_row(report)
top_nccl = _top_nccl_row(report)
if total_gpu_ms > 0.0 and top_nccl:
nccl_pct = (_coerce_float(top_nccl.get("total_time_ms")) / total_gpu_ms) * 100.0
kernel_pct = _coerce_float(top_kernel.get("pct_total_kernel_time") if top_kernel else 0.0)
if nccl_pct >= kernel_pct:
return "{} dominates {:.1f}% of GPU time".format(str(top_nccl.get("op_name") or "NCCL"), nccl_pct)
if top_kernel:
return "{} dominates {:.1f}% of GPU time".format(
str(top_kernel.get("kernel_name") or "Top kernel"),
_coerce_float(top_kernel.get("pct_total_kernel_time")),
)
return "No dominant GPU bottleneck detected from available metrics"
def _summary_rows(report: Mapping[str, Any]) -> List[Dict[str, str]]:
metrics = report.get("metrics") or {}
timeline = metrics.get("timeline") or {}
gpu_total_ms = _coerce_float(timeline.get("total_gpu_time_ms"))
if gpu_total_ms <= 0:
gpu_total_ms = _coerce_float((metrics.get("top_kernels") or {}).get("total_kernel_time_ns")) / 1_000_000.0
cpu_total_ms = _coerce_float(timeline.get("total_cpu_time_ms"))
if cpu_total_ms <= 0:
sync_rows = (metrics.get("sync") or {}).get("sync_calls") or []
cpu_total_ms = sum(_coerce_float(row.get("total_time_ms")) for row in sync_rows)
warnings = report.get("warnings") or []
report_version = _safe_text((report.get("tool") or {}).get("version"), default="unknown")
top_kernel = _top_kernel_row(report)
top_nccl = _top_nccl_row(report)
nvlink = (metrics.get("nvlink_during_nccl") or {}).get("rows") or []
nvlink_row = nvlink[0] if nvlink else None
capability_checks = {
"Kernel table": bool((metrics.get("top_kernels") or {}).get("present")),
"Runtime table": bool((metrics.get("sync") or {}).get("present")),
"NVTX ranges": bool((metrics.get("nvtx") or {}).get("present")),
"GPU metrics": bool((metrics.get("nvlink_during_nccl") or {}).get("present")),
"Per-process breakdown": bool((metrics.get("per_pid") or {}).get("present")),
}
rows: List[Dict[str, str]] = [
{"section": "Overview", "metric": "Trace", "value": _safe_trace_name(report)},
{"section": "Overview", "metric": "Tool version", "value": report_version},
{"section": "Overview", "metric": "Generated at (UTC)", "value": _safe_text(report.get("generated_at"))},
{"section": "Overview", "metric": "Total GPU time", "value": _format_ms(gpu_total_ms)},
{"section": "Overview", "metric": "Total CPU time", "value": _format_ms(cpu_total_ms)},
{"section": "Overview", "metric": "Top bottleneck", "value": _bottleneck_sentence(report)},
{"section": "Overview", "metric": "Warnings", "value": str(len(warnings))},
]
if top_kernel:
rows.extend(
[
{"section": "Evidence", "metric": "Top kernel", "value": _safe_text(top_kernel.get("kernel_name"))},
{"section": "Evidence", "metric": "Top kernel time", "value": _format_ms(top_kernel.get("total_time_ms"))},
{"section": "Evidence", "metric": "Top kernel share", "value": _format_pct(top_kernel.get("pct_total_kernel_time"))},
]
)
if top_nccl:
rows.extend(
[
{"section": "Evidence", "metric": "Top NCCL op", "value": _safe_text(top_nccl.get("op_name"))},
{"section": "Evidence", "metric": "Top NCCL time", "value": _format_ms(top_nccl.get("total_time_ms"))},
{"section": "Evidence", "metric": "Top NCCL overlap", "value": _format_pct(top_nccl.get("compute_overlap_pct"))},
]
)
if nvlink_row:
rows.extend(
[
{"section": "Evidence", "metric": "NVLink metric(s)", "value": _safe_text(nvlink_row.get("metric_names"))},
{
"section": "Evidence",
"metric": "NVLink during NCCL",
"value": "{:.2f} export units".format(_coerce_float(nvlink_row.get("avg_metric_during_nccl"), 0.0)),
},
{
"section": "Evidence",
"metric": "NVLink outside NCCL",
"value": "{:.2f} export units".format(_coerce_float(nvlink_row.get("avg_metric_outside_nccl"), 0.0)),
},
{
"section": "Evidence",
"metric": "NVLink correlation",
"value": "{:.3f}".format(_coerce_float(nvlink_row.get("nccl_activity_correlation"), 0.0)),
},
]
)
for label, present in capability_checks.items():
rows.append({"section": "Capabilities", "metric": label, "value": "present" if present else "missing"})
return rows
def _findings_markdown(report: Mapping[str, Any]) -> str:
findings = report.get("findings") or []
warnings = report.get("warnings") or []
lines: List[str] = ["## What to do next", ""]
if not findings:
lines.append("No findings were generated for this trace.")
else:
for finding in findings:
severity = _safe_text(finding.get("severity"), default="unknown").upper()
title = _safe_text(finding.get("title"), default="Untitled finding")
lines.append("### [{}] {}".format(severity, title))
evidence = finding.get("evidence") or []
recommendations = finding.get("recommendation") or finding.get("recommendations") or []
if evidence:
lines.append("Evidence:")
for item in evidence:
lines.append("- {}".format(item))
if recommendations:
lines.append("Recommendation:")
if isinstance(recommendations, (list, tuple)):
for item in recommendations:
lines.append("- {}".format(item))
else:
lines.append("- {}".format(recommendations))
lines.append("")
if warnings:
lines.append("## Warnings")
lines.append("")
for warning in warnings:
lines.append("- {}".format(warning))
return "\n".join(lines).strip()
def _artifact_manifest(out_dir: Path) -> List[Dict[str, str]]:
purpose_map = {
"report.md": "Human-readable report",
"report.json": "Machine-readable report",
"kernels.csv": "Top kernels",
"barriers.csv": "CPU/GPU barriers",
"nccl_ops.csv": "Top NCCL ops",
"nccl_rank_skew.csv": "Per-rank NCCL skew",
"nccl_by_pid.csv": "NCCL per PID",
"nvlink_during_nccl.csv": "NVLink correlation rows",
"nvlink_timeseries.csv": "NVLink correlation timeseries",
"timeline_events.csv": "Timeline events",
"copy_engine_events.csv": "Copy engine events",
"launch_latency_rows.csv": "Launch latency rows",
"launch_latency_histogram.csv": "Launch latency histogram",
"stream_overlap.csv": "Stream overlap summary",
"phase_split.csv": "Phase split",
"roofline.csv": "Roofline rows",
"gpu_idle_gaps.csv": "GPU idle gaps",
"kernels_by_pid.csv": "Per-PID kernels",
"sync_by_pid.csv": "Per-PID sync calls",
"nvtx_by_pid.csv": "Per-PID NVTX ranges",
"nvtx_ranges.csv": "NVTX ranges",
"bundle.zip": "Download all artifacts as a zip",
}
rows: List[Dict[str, str]] = []
for name, purpose in purpose_map.items():
path = out_dir / name
if not path.exists():
path = out_dir / "tables" / name
if path.exists():
rows.append({"artifact": name, "purpose": purpose, "path": str(path)})
return rows
def _zip_artifacts(out_dir: Path) -> Path:
zip_path = out_dir / "bundle.zip"
with zipfile.ZipFile(zip_path, "w", compression=zipfile.ZIP_DEFLATED) as zf:
for path in sorted(out_dir.rglob("*")):
if path.is_file() and path != zip_path:
zf.write(path, arcname=path.relative_to(out_dir).as_posix())
return zip_path
def _normalize_report_for_artifacts(report: Mapping[str, Any]) -> Dict[str, Any]:
normalized: Dict[str, Any] = dict(report)
metrics: Dict[str, Any] = dict(normalized.get("metrics") or {})
metrics.setdefault("top_kernels", {"present": False, "kernels": []})
metrics.setdefault("barriers", {"present": False, "barriers": []})
metrics.setdefault("nccl", {"present": False, "ops": [], "rank_rows": [], "pids": []})
metrics.setdefault("nvlink_during_nccl", {"present": False, "rows": [], "timeseries": []})
metrics.setdefault("timeline", {"present": False, "events": []})
metrics.setdefault("copy_engine", {"present": False, "events": []})
metrics.setdefault("launch_latency", {"present": False, "rows": [], "histogram": []})
metrics.setdefault("stream_overlap", {"present": False, "summary": []})
metrics.setdefault("phase_split", {"present": False, "rows": []})
metrics.setdefault("roofline", {"present": False, "rows": []})
metrics.setdefault("gpu_idle", {"present": False, "gaps": []})
metrics.setdefault("nvtx", {"present": False, "ranges": []})
by_pid = dict(metrics.get("by_pid") or {})
by_pid.setdefault("kernels", {"kernels": []})
by_pid.setdefault("sync", {"sync_calls": []})
by_pid.setdefault("nvtx", {"present": False, "ranges": []})
metrics["by_pid"] = by_pid
normalized["metrics"] = metrics
return normalized
def _load_report(path: Path) -> Tuple[str, Dict[str, Any], str]:
lower = path.suffix.lower()
if lower in (".sqlite", ".db"):
db = TraceDB.open(path)
try:
outputs = analyze(
db,
phase_map_path=None,
kernel_limit=50,
compute_kernel_percentiles=True,
compute_nvtx_kernel_map=True,
)
return "sqlite", dict(outputs.report), str(outputs.markdown)
finally:
db.close()
if lower == ".json":
report = json.loads(path.read_text(encoding="utf-8"))
if not isinstance(report, dict):
raise ValueError("Input JSON root must be an object.")
try:
markdown = render_markdown(report)
except Exception:
markdown = "# Nsight Systems LLM Hotspot Report\n\nJSON loaded, but markdown rendering failed for this input."
return "json", report, markdown
header = path.read_bytes()[:32]
if header.startswith(b"SQLite format 3"):
db = TraceDB.open(path)
try:
outputs = analyze(
db,
phase_map_path=None,
kernel_limit=50,
compute_kernel_percentiles=True,
compute_nvtx_kernel_map=True,
)
return "sqlite", dict(outputs.report), str(outputs.markdown)
finally:
db.close()
report = json.loads(path.read_text(encoding="utf-8"))
if not isinstance(report, dict):
raise ValueError("Input JSON root must be an object.")
try:
markdown = render_markdown(report)
except Exception:
markdown = "# Nsight Systems LLM Hotspot Report\n\nJSON loaded, but markdown rendering failed for this input."
return "json", report, markdown
def analyze_path(path: Path) -> SpaceBundle:
source_kind, report, markdown = _load_report(path)
report = _normalize_report_for_artifacts(report)
outputs = AnalysisOutputs(report=report, markdown=markdown)
artifacts_dir = Path(tempfile.mkdtemp(prefix="nsys-llm-explainer-space-")) / path.stem
write_artifacts(outputs, artifacts_dir)
_zip_artifacts(artifacts_dir)
artifact_paths = sorted(
[p for p in artifacts_dir.rglob("*") if p.is_file()],
key=lambda item: item.relative_to(artifacts_dir).as_posix(),
)
return SpaceBundle(
source_path=path,
source_kind=source_kind,
report=report,
markdown=markdown,
artifacts_dir=artifacts_dir,
artifact_paths=artifact_paths,
summary_rows=_summary_rows(report),
manifest_rows=_artifact_manifest(artifacts_dir),
findings_markdown=_findings_markdown(report),
status_markdown="Loaded `{}` as `{}` and wrote artifacts to `{}`.".format(path.name, source_kind, artifacts_dir),
)
def find_local_sample() -> Optional[Path]:
here = Path(__file__).resolve()
candidates = [here.parent / "sample_report.json"]
for parent in (here.parent, *tuple(here.parents)):
candidates.append(parent / "examples" / "synthetic" / "report.json")
candidates.append(parent / "examples" / "a100_vllm" / "report.json")
for candidate in candidates:
if candidate.exists():
return candidate
return None
def coerce_upload_path(uploaded: Any) -> Optional[Path]:
if uploaded is None:
return None
if isinstance(uploaded, (str, Path)):
path = Path(uploaded)
return path if path.exists() else None
if isinstance(uploaded, Sequence) and uploaded:
first = uploaded[0]
if isinstance(first, (str, Path)):
path = Path(first)
return path if path.exists() else None
return None
|