Qualify scoped Spark Omni text path on native Windows RTX
This commit is contained in:
1 parent
b4179a9c1c
commit
605b68cc5f
12 files changed
+2359
-11
No files matched your search
@@ -26,21 +26,25 @@ def clean_ledger(value: dict) -> bool:
|
||||
|
||||
|
||||
def audit(root: Path, route: str) -> dict:
|
||||
report_path = root / f"{route}_v4_report.json"
|
||||
first_log = root / f"{route}_v4_server.log"
|
||||
second_log = root / f"{route}_v4_server_restart.log"
|
||||
prefix = "spark_rtx" if route == "rtx" else f"{route}_v4"
|
||||
report_path = root / f"{prefix}_report.json"
|
||||
first_log = root / f"{prefix}_server.log"
|
||||
second_log = root / f"{prefix}_server_restart.log"
|
||||
run = json.loads(report_path.read_text(encoding="utf-8"))
|
||||
before = run["execution_plan"]
|
||||
after = run["restart_check"]["fresh_execution_plan"]
|
||||
abort = run["abort_check"]
|
||||
restart = run["restart_check"]
|
||||
if run["status"] != "passed" or run["device"] != ("cpu" if route == "cpu" else "Vulkan1"):
|
||||
expected_device = {"cpu": "cpu", "radeon": "Vulkan1", "rtx": "Vulkan0"}[route]
|
||||
if run["status"] != "passed" or run["device"] != expected_device:
|
||||
raise ValueError(f"{route}: runtime outcome or device differs")
|
||||
if route == "rtx" and (run["warmup_count"] != 1 or run["measured_count"] != 20):
|
||||
raise ValueError("RTX route did not complete its declared 20-request profile")
|
||||
if any(plan["model_sha256"] != MODEL_SHA256 or plan["server_sha256"] != SERVER_SHA256
|
||||
for plan in (before, after)):
|
||||
raise ValueError(f"{route}: model or server binary changed between sessions")
|
||||
expected_offload = ["0", "29"] if route == "cpu" else ["29", "29"]
|
||||
expected_buffers = [] if route == "cpu" else ["Vulkan1"]
|
||||
expected_buffers = {"cpu": [], "radeon": ["Vulkan1"], "rtx": ["Vulkan0"]}[route]
|
||||
if any(plan["offloaded_layers"] != expected_offload
|
||||
or plan["model_buffer_devices"] != expected_buffers for plan in (before, after)):
|
||||
raise ValueError(f"{route}: execution placement differs from the intended hardware")
|
||||
@@ -50,11 +54,17 @@ def audit(root: Path, route: str) -> dict:
|
||||
if route == "radeon" and any(plan["expected_device_name"] != "AMD Radeon(TM) 890M Graphics"
|
||||
for plan in (before, after)):
|
||||
raise ValueError("Radeon route did not pin the 890M by name")
|
||||
if route == "rtx" and any(plan["expected_device_name"] != "NVIDIA GeForce RTX 5090 Laptop GPU"
|
||||
for plan in (before, after)):
|
||||
raise ValueError("RTX route did not pin the 5090 by name")
|
||||
if ([row["answer"] for row in run["checks"]] != ["Paris", "1511"]
|
||||
or not all(row["correct"] for row in run["checks"] + run["measured"])
|
||||
or restart["fresh_request"]["answer"] != "Paris"
|
||||
or not restart["fresh_request"]["correct"]):
|
||||
raise ValueError(f"{route}: named complete-request checks failed")
|
||||
if route == "rtx" and (not run["context_check"]["rejected"]
|
||||
or run["context_check"]["recovery"]["answer"] != "Paris"):
|
||||
raise ValueError("RTX route did not reject overflow and recover")
|
||||
if (abort["abort_case"] != "inventory_120"
|
||||
or abort["server_tasks_after_abort_request_start"]
|
||||
<= abort["server_tasks_before_abort_request"]
|
||||
@@ -74,6 +84,28 @@ def audit(root: Path, route: str) -> dict:
|
||||
second_starts = len(TASK_START.findall(second_log.read_text(encoding="utf-8", errors="replace")))
|
||||
if first_starts < abort["server_tasks_after_abort_request_start"] or second_starts < 1:
|
||||
raise ValueError(f"{route}: server logs do not contain the recorded tasks")
|
||||
if route == "rtx":
|
||||
for path in (first_log, second_log):
|
||||
log = path.read_text(encoding="utf-8", errors="replace")
|
||||
if ("using device Vulkan0 (NVIDIA GeForce RTX 5090 Laptop GPU)" not in log
|
||||
or "offloaded 29/29 layers to GPU" not in log
|
||||
or "Vulkan0 model buffer size" not in log):
|
||||
raise ValueError("RTX raw server log does not verify GPU placement")
|
||||
public_path = root / "spark_rtx_public_report.json"
|
||||
public_log_path = root / "spark_rtx_public_server.log"
|
||||
public = json.loads(public_path.read_text(encoding="utf-8"))
|
||||
if (public["status"] != "passed" or public["device"] != "Vulkan0"
|
||||
or len(public["outputs"]) != 1
|
||||
or public["outputs"][0]["text"].strip() != "Paris"
|
||||
or public["outputs"][0]["error"] is not None):
|
||||
raise ValueError("RTX public AsyncOmni request did not pass")
|
||||
public_log = public_log_path.read_text(encoding="utf-8", errors="replace")
|
||||
if ("using device Vulkan0 (NVIDIA GeForce RTX 5090 Laptop GPU)" not in public_log
|
||||
or "offloaded 29/29 layers to GPU" not in public_log):
|
||||
raise ValueError("RTX public request did not use the RTX")
|
||||
files = [report_path, first_log, second_log]
|
||||
if route == "rtx":
|
||||
files += [public_path, public_log_path, root / "hardware.json"]
|
||||
return {
|
||||
"status": "passed",
|
||||
"device": run["device"],
|
||||
@@ -85,20 +117,22 @@ def audit(root: Path, route: str) -> dict:
|
||||
"fresh_answer": restart["fresh_request"]["answer"],
|
||||
"ledger_cleared": True,
|
||||
"files": {p.name: {"bytes": p.stat().st_size, "sha256": sha256(p)}
|
||||
for p in (report_path, first_log, second_log)},
|
||||
for p in files},
|
||||
}
|
||||
|
||||
|
||||
def main() -> None:
|
||||
parser = argparse.ArgumentParser(description=__doc__)
|
||||
parser.add_argument("--evidence-dir", required=True, type=Path)
|
||||
parser.add_argument("--routes", nargs="+", choices=("cpu", "radeon", "rtx"),
|
||||
default=("cpu", "radeon"))
|
||||
args = parser.parse_args()
|
||||
result = {
|
||||
"scope": "one in-flight cancel and fresh-stage restart per native Windows Spark placement",
|
||||
"scope": "one in-flight cancel and fresh-stage restart per selected native Windows Spark placement",
|
||||
"checkpoint": "Spark-X2.5-1.7B Q4_K_M",
|
||||
"model_sha256": MODEL_SHA256,
|
||||
"llama_server_sha256": SERVER_SHA256,
|
||||
"routes": {route: audit(args.evidence_dir, route) for route in ("cpu", "radeon")},
|
||||
"routes": {route: audit(args.evidence_dir, route) for route in args.routes},
|
||||
"limits": [
|
||||
"A fresh StageRuntime is created after cancellation; no same-session state migration is claimed.",
|
||||
"The backend returns complete requests rather than incremental token events.",
|
||||
|
||||
Reference in new issue
Block a user