print log about runtime cocurrency

lilyscript/generator.py

@@ -15,6 +15,9 @@ log) and `pretty = postprocess(raw)` (for the editor, segmented by measure).
 
 import os
 import json
+import logging
+
+LOG = logging.getLogger('lilyscript')
 
 import numpy as np
 import onnxruntime as ort
@@ -68,6 +71,44 @@ def _softmax (x):
 	return e / e.sum()
 
 
+def _physical_cores ():
+	'''Best-effort physical (not logical/HT) core count via /proc/cpuinfo; None if
+	unavailable. ORT's intra_op default (=0) maps to this on most CPU builds.'''
+	try:
+		phys, cur = set(), {}
+		for line in open('/proc/cpuinfo'):
+			line = line.strip()
+			if not line:
+				if 'physical id' in cur and 'core id' in cur:
+					phys.add((cur['physical id'], cur['core id']))
+				cur = {}
+				continue
+			if ':' in line:
+				k, v = line.split(':', 1)
+				cur[k.strip()] = v.strip()
+		return len(phys) or None
+	except Exception:
+		return None
+
+
+def _log_thread_info (so, sess):
+	'''Log host CPU capacity + the ONNX Runtime intra/inter-op thread settings that
+	are actually in effect. intra_op_num_threads/inter_op_num_threads == 0 means
+	"ORT auto" — it picks the number of physical cores for the intra-op pool.'''
+	logical = os.cpu_count()
+	affinity = len(os.sched_getaffinity(0)) if hasattr(os, 'sched_getaffinity') else logical
+	physical = _physical_cores()
+	intra = so.intra_op_num_threads
+	inter = so.inter_op_num_threads
+	effective_intra = intra if intra else (physical or affinity or logical)
+	LOG.info('CPU: %s logical / %s physical cores, %s available (affinity)',
+		logical, physical if physical is not None else '?', affinity)
+	LOG.info('ONNX Runtime threads: intra_op=%s (%s), inter_op=%s (%s) | execution_mode=%s',
+		intra, 'auto -> ~%s' % effective_intra if intra == 0 else 'explicit',
+		inter, 'auto' if inter == 0 else 'explicit',
+		getattr(so, 'execution_mode', '?'))
+
+
 class StreamingLilyletGenerator:
 	'''Loads the int8 KV ONNX sessions + vendored assets and streams generation.'''
 
@@ -100,6 +141,7 @@ class StreamingLilyletGenerator:
 			os.path.join(model_dir, 'token_kv_int8.onnx'), so, providers=['CPUExecutionProvider'])
 		self.patch_out_names = [o.name for o in self.patch_kv_sess.get_outputs()]
 		self.token_out_names = [o.name for o in self.token_kv_sess.get_outputs()]
+		_log_thread_info(so, self.patch_kv_sess)
 
 	# ---- text helpers (mirror LilyletPatchyGenerator.patch_to_text) ----