fix(kyojin): reuse queued prefix pages across agent turns

Bead: nix-overlay-7p6
This commit is contained in:
2026-11-19 01:28:42 +03:00
parent 7843d79e76
commit 635f9f7dfc
3 changed files with 186 additions and 0 deletions
+1
View File
@@ -9,6 +9,7 @@ let
enginePatches = [
./prefix-disk-admission.patch
./prefix-writer-release.patch
./prefix-inflight-pages.patch
./expert-direct-read-eof.patch
];
engine = exllamav3.overrideAttrs (old: {
+184
View File
@@ -0,0 +1,184 @@
--- a/exllamav3/generator/prefix_disk.py
+++ b/exllamav3/generator/prefix_disk.py
@@ -23,6 +23,7 @@
import hashlib
import os
+import threading
import time
import torch
@@ -90,6 +91,7 @@
self.pages: dict[bytes, tuple[str, int]] = {} # page hash -> (segment, position in segment)
self.stashes: dict[bytes, tuple[str, int]] = {} # anchor hash -> (entry, depth in pages)
self.tails: dict[str, dict] = {} # entry -> idx
+ self._segment_writes: list[tuple[threading.Event, list[bytes]]] = []
self.pending: list[torch.Tensor] = []
self._staging = ([None, None], [None, None]) # pinned host buffers + their last device-copy events
self.stats = {"saves": 0, "save_ms": 0.0, "restores": 0, "restore_ms": 0.0, "restored_tokens": 0}
@@ -130,6 +132,7 @@
def clear(self) -> None:
"""Forget everything (cache_salt change): the entries were computed under another namespace."""
self.disk.flush()
+ self._segment_writes.clear()
for name in self.disk.entries():
self.disk.remove(name)
self.pages, self.stashes, self.tails = {}, {}, {}
@@ -185,7 +188,11 @@
if ent[0] not in ok:
ok[ent[0]] = self._have(ent[0])
return ok[ent[0]]
- new = [i for i in range(keep_depth) if not on_disk(hashes[i])]
+ # A follow-up can finish before the previous snapshot publishes its index.
+ # Save can reuse these exact host copies. Restore requires a published index.
+ self._segment_writes = [(done, hs) for done, hs in self._segment_writes if not done.is_set()]
+ staged = {h for _, hs in self._segment_writes for h in hs}
+ new = [i for i in range(keep_depth) if hashes[i] not in staged and not on_disk(hashes[i])]
wrote = 0
if new:
src = [pages[i] for i in new]
@@ -201,8 +208,10 @@
tens.append(t.index_select(0, idx).cpu())
hs = [hashes[i] for i in new]
name = f"s_{_hx(hs[0])[:20]}_{len(hs)}"
- self.disk.put(name, {"hashes": hs, "sequences": [p.sequence.cpu().clone() for p in src], "tensors": tens},
- {"ns": self.ns, "k": "seg", "hashes": hs}, min_bytes=min_bytes)
+ done = self.disk.put(name, {"hashes": hs, "sequences": [p.sequence.cpu().clone() for p in src], "tensors": tens},
+ {"ns": self.ns, "k": "seg", "hashes": hs}, min_bytes=min_bytes)
+ if done is not None:
+ self._segment_writes.append((done, hs))
for pos, h in enumerate(hs):
self.pages[h] = (name, pos)
wrote += len(hs)
--- a/exllamav3/util/disk_store.py
+++ b/exllamav3/util/disk_store.py
@@ -161,10 +161,12 @@
return False
return True
- def put(self, name: str, payload, idx, *, min_bytes: int = 0) -> None:
- """Queue an entry: payload and idx are torch.save-able and must not change afterwards (host copies)."""
+ def put(self, name: str, payload, idx, *, min_bytes: int = 0) -> threading.Event | None:
+ """Queue immutable host copies. For byte-budgeted segments, signal when the write attempt ends."""
if not self._closed:
- self._q.put((name, payload, idx, min_bytes))
+ finished = threading.Event() if min_bytes else None
+ self._q.put((name, payload, idx, min_bytes, finished))
+ return finished
def flush(self) -> None:
self._q.join()
@@ -180,11 +182,12 @@
def _run(self) -> None:
window_start, window_bytes = time.monotonic(), 0
while True:
+ finished = None
item = self._q.get()
try:
if item is None:
return
- name, payload, idx, min_bytes = item
+ name, payload, idx, min_bytes, finished = item
if os.path.exists(self._p(name, ".idx")):
self.touch(name)
continue
@@ -215,6 +218,9 @@
spent = time.monotonic() - window_start
ahead = window_bytes / self.rate - spent # stay under the per-minute write budget
if ahead > 0:
+ if finished is not None:
+ finished.set()
+ finished = None
item = payload = idx = None # Release host tensors before write-rate throttling.
time.sleep(min(ahead, 120.0))
if spent > 60:
@@ -228,4 +234,6 @@
pass
finally:
item = payload = idx = None # Skipped and failed writes can also precede an idle queue.
+ if finished is not None:
+ finished.set()
self._q.task_done()
--- a/tests/test_disk_store_cpu.py
+++ b/tests/test_disk_store_cpu.py
@@ -145,8 +145,9 @@
try:
assert s.can_stage(n)
free.return_value = SimpleNamespace(free=s.floor + n)
- s.put("seg", payload, {}, min_bytes=n)
+ finished = s.put("seg", payload, {}, min_bytes=n)
s.flush()
+ assert finished.is_set()
save.assert_not_called()
assert not s.entries() and s.stats["skipped_floor"] == 1
finally:
@@ -165,9 +166,10 @@
tensor = torch.arange(32)
ref = weakref.ref(tensor)
try:
- store.put("seg", {"x": tensor}, {})
+ finished = store.put("seg", {"x": tensor}, {}, min_bytes=tensor.nbytes)
del tensor
assert sleeping.wait(5)
+ assert finished.is_set()
assert ref() is None
assert torch.equal(store.read("seg")["x"], torch.arange(32))
finally:
@@ -187,3 +189,57 @@
assert ref() is None
finally:
store.close()
+
+
+@pytest.mark.parametrize("fail_first", [False, True])
+def test_consecutive_saves_share_only_inflight_pages(tmp_path, monkeypatch, fail_first):
+ from exllamav3.constants import PAGE_SIZE
+ from exllamav3.generator.prefix_disk import PagedPrefixDisk
+
+ ids = torch.arange(2 * PAGE_SIZE).view(1, -1)
+ pool = torch.arange(6 * 128).view(6, -1)
+ pages = [SimpleNamespace(page_index=i, sequence=ids[:, n * PAGE_SIZE:(n + 1) * PAGE_SIZE])
+ for n, i in enumerate((4, 2))]
+ hashes = PagedPrefixDisk.chain(ids)
+ rc = Mock()
+ rc.get.side_effect = lambda h: {"position": (1 + hashes.index(h)) * PAGE_SIZE}
+ def chain(tokens):
+ count = tokens.numel() // PAGE_SIZE
+ return hashes[:count], pages[:count]
+ pages_on_device = SimpleNamespace(referenced_pages={}, unreferenced_pages={})
+ generator = SimpleNamespace(recurrent_cache=rc, pagetable=pages_on_device, tail_ckpts=[])
+ slots = SimpleNamespace(generator=generator, _tensors=lambda: [pool],
+ _resident_chain=chain, _find_tail=lambda *_: None)
+ pd = PagedPrefixDisk(slots, str(tmp_path))
+ pd.disk.floor = 0
+ pd.disk.rate = 1e12
+ monkeypatch.setattr(torch.cuda, "synchronize", lambda: None)
+ started, resume = Event(), Event()
+ serialize = torch.save
+ def hold_first(obj, path, *args, **kwargs):
+ if isinstance(obj, dict) and obj.get("hashes") == hashes[:1] and not started.is_set():
+ started.set()
+ assert resume.wait(10)
+ if fail_first:
+ raise OSError("injected write failure")
+ return serialize(obj, path, *args, **kwargs)
+ monkeypatch.setattr(torch, "save", hold_first)
+ try:
+ assert pd.save(ids[:, :PAGE_SIZE])["saved"] == 1
+ assert started.wait(5)
+ assert not pd._have(pd.pages[hashes[0]][0])
+ assert pd.restore(ids) is None
+ assert pd.save(ids)["saved"] == 1
+ resume.set()
+ pd.disk.flush()
+ assert pd.save(ids)["saved"] == int(fail_first)
+ pd.disk.flush()
+ for h, page in zip(hashes, pages):
+ name, pos = pd.pages[h]
+ assert torch.equal(pd.disk.read(name)["tensors"][0][pos], pool[page.page_index])
+ # A committed segment that is later evicted must be staged again.
+ pd.disk.remove(pd.pages[hashes[0]][0])
+ assert pd.save(ids)["saved"] == 1
+ finally:
+ resume.set()
+ pd.disk.close()