Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.
C++
launch: a GPU runtime start failure (hsa_exception) says the card was not ready, with a next step; the runtime text stays in the log
Niko122160d847ba3389
diff --git a/strata_cli/launch/runner.py b/strata_cli/launch/runner.pyindex b408ff739..abe7e5759 100644--- a/strata_cli/launch/runner.py+++ b/strata_cli/launch/runner.py@@ -1967,6 +1967,11 @@ def _read_json(path: Path) -> dict | None:#: lines of an engine log that say why it stopped (the last one wins), and a plain next step for the causes we know. Recommend, never force.CAUSE_LINE = re.compile(r"error|exception|failed|cannot|can't|could not|no device|not found|out of memory|unsupported|abort|invalid|missing|"r"no kernel|illegal|denied|refus", re.I)+# The GPU runtime (ROCm) failed while the card was being set up, often right after another engine exited: the runtime's own text+# ("terminate called after throwing an instance of 'rocr::AMD::hsa_exception'") stays in the log, not in the sentence shown.+HSA_START = re.compile(r"hsa_exception|HSA_STATUS_ERROR|hsaKmt", re.I)+HSA_START_NEXT = ("The graphics card was not ready yet; wait a few seconds and run it again. If it keeps happening, restart the PC "+ "and report the log.")KNOWN_CAUSES = ((re.compile(r"no device of requested type|device not found|no (gpu|sycl|level.zero) device|permission denied.*(dri|render)", re.I),"The engine could not open the graphics card. On Linux the account must be allowed to use it: add it to the render group "@@ -2045,10 +2050,13 @@ def explain_engine_exit(result: dict, config: Path | None, marks: dict | None =stage = "The engine stopped while it was loading the model" + (f" (exit code {code})" if code not in (None, 0) else "")else:stage = f"The server did not become ready ({state}: {why or 'see the engine log'})"+ hsa = HSA_START.search(" ".join(lines + [str(why or "")])) # the runtime's own text is kept out of the sentence (it is in the log)+ if hsa:+ cause = ""text = stage + (f": {cause}" if cause else ".")if not text.endswith((".", "!")):text += "."- step = next((words for pattern, words in KNOWN_CAUSES if pattern.search(" ".join([cause, str(why or "")]))), "Run it again; if it repeats, report the log.")+ step = HSA_START_NEXT if hsa else next((words for pattern, words in KNOWN_CAUSES if pattern.search(" ".join([cause, str(why or "")]))), "Run it again; if it repeats, report the log.")return text, step + (f" The engine's log: {log}" if log else " The engine's log has the reason (strata logs, or the log named in the run config).")diff --git a/tests/terminal/test_round8_fixes.py b/tests/terminal/test_round8_fixes.pyindex 1a8914832..47ef02764 100644--- a/tests/terminal/test_round8_fixes.py+++ b/tests/terminal/test_round8_fixes.py@@ -44,6 +44,22 @@ class EngineExitIsExplained(unittest.TestCase):self.assertIn("render group", nxt)self.assertIn(str(log), nxt)+ def test_a_gpu_runtime_start_failure_gets_a_plain_sentence_and_keeps_the_runtime_text_out_of_it(self):+ cfg, log = self.config_with_log(["strata generate: loading", "terminate called after throwing an instance of 'rocr::AMD::hsa_exception'",+ " what(): "])+ text, nxt = R.explain_engine_exit({"state": "failed", "error": "server_exited", "engine_exit_code": 134}, cfg)+ self.assertIn("stopped while it was loading the model (exit code 134)", text)+ self.assertNotIn("hsa_exception", text + nxt)+ self.assertNotIn("terminate called", text + nxt)+ self.assertIn("graphics card was not ready yet", nxt)+ self.assertIn("wait a few seconds", nxt)+ self.assertIn("restart the PC", nxt)+ self.assertIn(str(log), nxt)+ text, nxt = R.explain_engine_exit({"error": "server_exited", "engine_exit_code": 134,+ "log_tail": "terminate called after throwing an instance of 'rocr::AMD::hsa_exception' | what():"}, None)+ self.assertNotIn("hsa_exception", text)+ self.assertIn("graphics card was not ready yet", nxt)+def test_out_of_memory_and_old_engine_get_their_own_next_step(self):cfg, _ = self.config_with_log(["strata generate: CUDA error: out of memory"])self.assertIn("--context 16k", R.explain_engine_exit({"error": "server_exited"}, cfg)[1])