Skip to content

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, ASR, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.

updates

HenryNdubuakue7a5dd8eb233
diff --git a/src/cli.py b/src/cli.py
index 854fee3..2299092 100644
--- a/src/cli.py
+++ b/src/cli.py
@@ -1,11 +1,100 @@
import argparse
import os
+import re
import sys
+import threading
from .data import DEFAULT_MAX_ENC_LEN, DEFAULT_MAX_DEC_LEN, DEFAULT_MAX_GEN_LEN
HELP = """Check the readme"""
+
+_ABSL_LOG_START = re.compile(rb"^[EIWF]\d{4} \d\d:\d\d:\d\d")
+_NOISY_LOG_HEADER = re.compile(
+ rb"\] (?:Fusion: .*gemm_fusion|Computation: .*_computation|Delay kernel timed out)"
+)
+
+_log_filter_installed = False
+
+
+def _install_xla_log_filter():
+ """Drop XLA Triton autotuner noise from stderr.
+
+ XLA's Triton GEMM autotuner logs failed candidate fusions via LOG(ERROR)
+ in xtile_compiler.cc and cuda_timer.cc. These are unconditional and do
+ not respect TF_CPP_MIN_LOG_LEVEL, so we filter them at the file
+ descriptor level.
+
+ Strategy:
+ - Rebind Python's sys.stderr to a fresh file object over the real
+ terminal fd, so tqdm and print() writes go straight to the terminal
+ and never enter our pipe. This keeps progress bars (which use \\r
+ without trailing \\n) from stalling the filter's line parser.
+ - Replace fd 2 with a pipe. Only C-level writes (absl / XLA LOG(...))
+ now flow through the pipe, and they are always \\n-terminated and
+ well-formed, so a simple line-based filter is reliable.
+ """
+ global _log_filter_installed
+ if _log_filter_installed:
+ return
+ _log_filter_installed = True
+
+ # 1. Dup the original stderr fd; rebind Python's sys.stderr to it so
+ # all Python-level writes bypass the pipe.
+ py_stderr_fd = os.dup(2)
+ try:
+ sys.stderr.flush()
+ except Exception:
+ pass
+ sys.stderr = os.fdopen(py_stderr_fd, "w", encoding="utf-8",
+ errors="replace", buffering=1)
+
+ # 2. Dup another copy for the filter thread to write filtered C-level
+ # output to.
+ out_fd = os.dup(2) # before dup2 below, fd 2 is still the real terminal
+
+ # 3. Redirect fd 2 to the write end of a new pipe. All C-level writes
+ # from XLA now land in the pipe.
+ r_fd, w_fd = os.pipe()
+ os.dup2(w_fd, 2)
+ os.close(w_fd)
+
+ def pump():
+ reader = os.fdopen(r_fd, "rb", buffering=0)
+ out = os.fdopen(out_fd, "wb", buffering=0)
+ buf = b""
+ skipping = False
+ try:
+ while True:
+ chunk = reader.read(65536)
+ if not chunk:
+ break
+ buf += chunk
+ while True:
+ idx = buf.find(b"\n")
+ if idx == -1:
+ break
+ line = bytes(buf[:idx])
+ buf = buf[idx + 1:]
+ is_log_start = bool(_ABSL_LOG_START.match(line))
+ if skipping:
+ if is_log_start:
+ if _NOISY_LOG_HEADER.search(line):
+ continue
+ skipping = False
+ out.write(line + b"\n")
+ # else: continuation body of a skipped log block — drop
+ else:
+ if is_log_start and _NOISY_LOG_HEADER.search(line):
+ skipping = True
+ continue
+ out.write(line + b"\n")
+ except Exception:
+ pass
+
+ t = threading.Thread(target=pump, daemon=True, name="xla-log-filter")
+ t.start()
+
def main():
if len(sys.argv) < 2 or sys.argv[1] in ("-h", "--help", "help"):
print(HELP)
@@ -245,12 +334,18 @@ def main():
from .tokenize_data import tokenize
tokenize(args)
elif args.command == "pretrain":
+ os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "2")
+ os.environ.setdefault("GRPC_VERBOSITY", "ERROR")
+ _install_xla_log_filter()
import jax
if os.path.exists("/dev/accel0"):
jax.distributed.initialize()
from .pretrain import pretrain
pretrain(args)
elif args.command == "train":
+ os.environ.setdefault("TF_CPP_MIN_LOG_LEVEL", "2")
+ os.environ.setdefault("GRPC_VERBOSITY", "ERROR")
+ _install_xla_log_filter()
import jax
if os.path.exists("/dev/accel0"):
jax.distributed.initialize()