[Bugfix][HiSparse] Fix MTP acceptance collapse under FULL graphs with a saturated GPU pool (#59309)

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Lucas Wilkinson
2026-10-01 23:04:21 -07:00
committed by GitHub
co-authored by Claude Opus 5.5 Matthew Bonanni
parent 0cbac6cd13
commit 4056c8ac1f
11 changed files with 426 additions and 122 deletions
+2 -24
View File
@@ -44,6 +44,8 @@ def _make_hisparse_worker() -> HiSparseConnectorWorker:
worker._row_mirror_num_rows = 0
worker._per_layer_mirrored = set()
worker._submitted_mirror_layers = set()
worker._draft_layers = ()
worker._draft_mirror_pending = False
worker._pending_dma_descriptors = deque()
worker._dma_free_descriptors = []
worker.host_write_events = (MagicMock(), MagicMock())
@@ -785,30 +787,6 @@ def test_hisparse_finish_forward_does_not_repeat_per_layer_mirrors():
worker._enqueue_row_dma.assert_not_called()
def test_hisparse_finish_forward_submits_lazy_post_forward_transfer(monkeypatch):
runtime = SimpleNamespace(eager_host_mirror=False)
worker = _make_hisparse_worker()
worker.is_host_writer = True
worker.cache_handles = [SimpleNamespace(runtime=runtime, num_actual_tokens=0)]
transfer = SparseKVPageTransfer(7, 2, (1,), after_forward=True)
worker._post_forward_transfers = [transfer]
worker._forward_ready_event = MagicMock()
worker._enqueue_host_mirror = MagicMock()
worker._submit_transfers = MagicMock()
worker._dma_submitted = False
worker._submitted_mirror_layers = set()
worker._finish_mirror_phase = MagicMock()
worker._release_completed_dma_descriptors = MagicMock()
stream = MagicMock()
monkeypatch.setattr(hisparse_worker_module, "current_stream", lambda: stream)
worker.finish_forward()
worker._finish_mirror_phase.assert_called_once_with(worker._forward_ready_event)
worker._submit_transfers.assert_called_once_with([transfer])
worker.host_write_event.record.assert_called_once_with(stream)
def test_hisparse_prefill_mirrors_source_groups_and_flushes_partial_group():
slots = torch.tensor([7, 8], dtype=torch.int64)
source_indices = [0, 0, 1, 1, 1, 1, 2]