fix: keep the refusal-direction projection dynamo-safe

With VLLM_ABLITERATION set, boot crashed: dynamo traces
abliteration.register from Qwen3NextGatedDeltaNet.__init__, and
buf.copy_(r) fails on the meta device during the compile-time model
construction pass. Disable tracing for register and guard the copy so
an unmaterializable buffer turns projection off with a warning.

Log activation at registration instead of inside project_: project_
runs in the compiled region, and a logger call there aborts dynamo.
The width guard uses the same escape hatch.

Also project the embedding that Qwen3NextMultiTokenPredictor builds
for itself, which the first pass missed, and add tests for the
2560-value Flash-Next direction against a 5120 model and for a
half-width activation.

Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
2026-10-15 22:56:06 +07:00
co-authored by Hinata
parent 0076adc78b
commit 5d6754ff6b
3 changed files with 75 additions and 14 deletions
+35
View File
@@ -16,8 +16,10 @@ HIDDEN = 8
@pytest.fixture(autouse=True)
def clear_direction_cache():
abliteration._read.cache_clear()
abliteration.clear_logs()
yield
abliteration._read.cache_clear()
abliteration.clear_logs()
def write_direction(tmp_path, name="r.json", size=HIDDEN):
@@ -97,6 +99,39 @@ def test_hidden_size_mismatch_raises(tmp_path, monkeypatch):
registered(torch.nn.Module(), size=HIDDEN + 1)
def test_first_projection_logs_that_it_is_active(tmp_path, monkeypatch, caplog):
"""The boot log proves the switch is live, not just loaded."""
import logging
monkeypatch.setenv("VLLM_ABLITERATION", write_direction(tmp_path))
module = registered(torch.nn.Linear(HIDDEN, HIDDEN))
with caplog.at_level(
logging.INFO, logger="vllm.model_executor.models.abliteration"
):
abliteration.project_(module, torch.randn(2, HIDDEN))
assert "active, projected a 8-wide tensor" in caplog.text
def test_flash_next_direction_is_rejected_by_the_27b_model(tmp_path, monkeypatch):
"""Flash-Next stores 2560 values; the dense 27B stream is 5120 wide."""
path = write_direction(tmp_path, name="flash_next.json", size=2560)
monkeypatch.setenv("VLLM_ABLITERATION", path)
with pytest.raises(
ValueError, match="holds 2560 values; the model's hidden size is 5120"
):
registered(torch.nn.Module(), size=5120)
def test_half_width_activation_is_skipped(tmp_path, monkeypatch):
"""A 2560-wide tensor cannot be projected by the 5120 direction."""
monkeypatch.setenv("VLLM_ABLITERATION", write_direction(tmp_path, size=5120))
module = registered(torch.nn.Module(), size=5120)
x = torch.randn(4, 2560)
before = x.clone()
abliteration.project_(module, x)
assert torch.equal(x, before)
def test_zero_direction_raises(tmp_path, monkeypatch):
path = tmp_path / "zero.json"
path.write_text(json.dumps({"refusal_direction": [0.0] * HIDDEN}))