Bounded truncation pool (8 threads); human eta up to days

95 waiters all tokenizing 2M-token docs through the 32-thread default
executor saturated the GIL: the rich render thread and the event loop
starved, so the bar froze and jumped (and the gate probe went blind).
Truncation now runs on a dedicated 8-thread pool; the remaining
workers queue and the loop/renderer stay responsive.

eta formats as 45s / 7m15s / 2h35m40s / 6d03h12m as it grows.

Co-Authored-By: Claude <noreply@anthropic.com>

evalharness/model/runner.py

@@ -319,11 +319,21 @@ async def generate_predictions(
 
         # assemble() tokenizes for the max_input_tokens truncation -- on
         # long-context benches that is SECONDS of CPU per sample (2M-token
-        # docs), and running it inline FROZE the whole event loop: heartbeat,
-        # gate probes and every other request serialized behind one encode.
-        # Thread it: the loop stays live and encodes parallelize (the Rust
-        # fast tokenizer releases the GIL).
-        text = await asyncio.to_thread(assemble, sample) \
+        # docs). Two failure modes fixed here:
+        #  - inline: froze the whole event loop behind one encode
+        #  - asyncio.to_thread (32-thread default pool): dozens of concurrent
+        #    tokenizers hogged the GIL and starved the progress renderer +
+        #    loop itself (bar froze, then jumped)
+        # A DEDICATED BOUNDED pool: 8 encodes at a time, remaining workers
+        # queue -- GIL pressure capped, everything stays responsive.
+        global _ASSEMBLE_EXEC
+        if _ASSEMBLE_EXEC is None:
+            import concurrent.futures
+
+            _ASSEMBLE_EXEC = concurrent.futures.ThreadPoolExecutor(
+                max_workers=8, thread_name_prefix='assemble')
+        text = await asyncio.get_running_loop().run_in_executor(
+            _ASSEMBLE_EXEC, assemble, sample) \
             if isinstance(sample.input, str) else None
         messages = ([ChatMessage(role='user', content=text)]
                     if isinstance(sample.input, str) else list(sample.input))
@@ -540,6 +550,7 @@ def _progress(progress: bool, done: int, total: int, t0: float, usage: Usage) ->
 
 
 _PROBED_SPECS = set()
+_ASSEMBLE_EXEC = None   # bounded truncation pool (lazy)
 
 
 async def _probe_model(adapter, model_spec: str) -> None:

evalharness/progress/rich_terminal.py

@@ -19,6 +19,21 @@ def _fmt(sec):
     return f'{sec // 60}m{sec % 60:02d}s' if sec >= 60 else f'{sec}s'
 
 
+def _fmt_eta(sec):
+    """eta: s -> m s -> h m s -> d h m (grows with the unit that matters)."""
+    sec = int(sec)
+    if sec < 60:
+        return f'{sec}s'
+    m, s = divmod(sec, 60)
+    if m < 60:
+        return f'{m}m{s:02d}s'
+    h, m = divmod(m, 60)
+    if h < 24:
+        return f'{h}h{m:02d}m{s:02d}s'
+    d, h = divmod(h, 24)
+    return f'{d}d{h:02d}h{m:02d}m'
+
+
 class RichTerminalProgress:
     def __init__(self, console=None):
         # accept an EXTERNAL console: CLI phase messages and the live bar must
@@ -168,7 +183,7 @@ class RichTerminalProgress:
             total=total, completed=min(done, total), new='',
             rate=f'{done / elapsed:.2f}', inflight=0, cur='0s',
             elapsed=_fmt(elapsed),
-            eta=_fmt((total - done) * elapsed / done) if done and total > done else '-')
+            eta=_fmt_eta((total - done) * elapsed / done) if done and total > done else '-')
 
     def set_bench_tag(self, tag: str):
         if self.disabled:
@@ -253,7 +268,7 @@ class RichTerminalProgress:
             rate=f"{fresh / elapsed:.2f}",
             inflight=self._inflight_txt(), cur='0s',
             elapsed=_fmt(elapsed),
-            eta=_fmt((task.total - completed) * elapsed / fresh)
+            eta=_fmt_eta((task.total - completed) * elapsed / fresh)
                  if fresh and task.total and task.total > completed else '-',
         )