Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Forward the model argument through update_weight_conversions overrides (transformers 5.21+) (#16559)

Daniel Hanca762a0ea9db
diff --git a/tests/test_compressed_tensors_bnb.py b/tests/test_compressed_tensors_bnb.py
index d0ad2eb2b84..4687c3f37ab 100644
--- a/tests/test_compressed_tensors_bnb.py
+++ b/tests/test_compressed_tensors_bnb.py
@@ -13,6 +13,7 @@
"""Packed INT4/INT8 checkpoints load into bnb 4-bit; GPU tests match a bf16 decompress bit for bit."""
+import inspect
import json
import os
import shutil
@@ -52,6 +53,14 @@ from unsloth.models.compressed_tensors_bnb import _transformers_supports_weight_
HAS_CONVERTERS = _transformers_supports_weight_converters()
+def _update_conversions(quantizer, conversions, model):
+ # transformers 5.21+ passes the model to the hook.
+ from transformers.quantizers.base import HfQuantizer
+ if "model" in inspect.signature(HfQuantizer.update_weight_conversions).parameters:
+ return quantizer.update_weight_conversions(conversions, model)
+ return quantizer.update_weight_conversions(conversions)
+
+
def _w4a16(**overrides):
weights = {
"num_bits": 4,
@@ -514,7 +523,7 @@ def test_the_planner_pass_does_not_consume_the_plan():
assert getattr(config, UNSLOTH_COMPRESSED_TENSORS_ATTR, None) is plan
second, model = preprocess() # the real load
assert second._unsloth_ct_config is not None
- conversions = second.update_weight_conversions([])
+ conversions = _update_conversions(second, [], model)
assert any(
getattr(c, "source_patterns", None) and "weight_packed$" in c.source_patterns
for c in conversions
@@ -542,8 +551,8 @@ def test_armed_conversions_still_run_a_composed_subclass_hook(monkeypatch):
marker = object()
class MarkerBnb4BitHfQuantizer(Bnb4BitHfQuantizer):
- def update_weight_conversions(self, weight_conversions):
- return super().update_weight_conversions(weight_conversions) + [marker]
+ def update_weight_conversions(self, weight_conversions, *args, **kwargs):
+ return super().update_weight_conversions(weight_conversions, *args, **kwargs) + [marker]
monkeypatch.setitem(
quantizers_auto.AUTO_QUANTIZER_MAPPING, "bitsandbytes_4bit", MarkerBnb4BitHfQuantizer
@@ -574,7 +583,7 @@ def test_armed_conversions_still_run_a_composed_subclass_hook(monkeypatch):
model, dtype = torch.bfloat16, device_map = None, checkpoint_files = []
)
assert quantizer._unsloth_ct_config is not None
- conversions = quantizer.update_weight_conversions([])
+ conversions = _update_conversions(quantizer, [], model)
assert marker in conversions
assert any(
getattr(c, "source_patterns", None) and "weight_packed$" in c.source_patterns
diff --git a/tests/test_k3_mxfp4_stacked_experts.py b/tests/test_k3_mxfp4_stacked_experts.py
index 5d45a69bee8..ed40e656b6e 100644
--- a/tests/test_k3_mxfp4_stacked_experts.py
+++ b/tests/test_k3_mxfp4_stacked_experts.py
@@ -13,10 +13,12 @@
"""Kimi-K3 style remote-code MXFP4 experts load packed into one Mxfp4StackedExperts per MoE layer."""
+import inspect
import json
import os
import sys
import types
+from types import SimpleNamespace
import re
@@ -804,10 +806,18 @@ def test_a_kept_packed_load_registers_no_decompress_op():
quantizer._unsloth_dtype_plan, quantizer._unsloth_plan_dicts = {}, []
ops = lambda convs: [type(op).__name__ for c in convs for op in getattr(c, "operations", [])] # noqa: E731
quantizer._unsloth_keep_packed, quantizer._unsloth_packed_experts = True, ["layers.0.mlp"]
- kept = ops(quantizer.update_weight_conversions([]))
+ from transformers.quantizers.base import HfQuantizer
+
+ # transformers 5.21+ passes the model to the hook.
+ extra = (
+ (SimpleNamespace(named_parameters = lambda: iter(())),)
+ if "model" in inspect.signature(HfQuantizer.update_weight_conversions).parameters
+ else ()
+ )
+ kept = ops(quantizer.update_weight_conversions([], *extra))
assert kept.count("StackPackedExperts") == 4 and "DecompressPackedWeights" not in kept
quantizer._unsloth_keep_packed, quantizer._unsloth_packed_experts = False, []
- assert "DecompressPackedWeights" in ops(quantizer.update_weight_conversions([]))
+ assert "DecompressPackedWeights" in ops(quantizer.update_weight_conversions([], *extra))
def _peft_target_parameters(model, monkeypatch, **kwargs):
diff --git a/tests/test_quantizer_update_weight_conversions_model.py b/tests/test_quantizer_update_weight_conversions_model.py
new file mode 100644
index 00000000000..1c673069b36
--- /dev/null
+++ b/tests/test_quantizer_update_weight_conversions_model.py
@@ -0,0 +1,96 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""transformers 5.21+ passes the model to update_weight_conversions. Unsloth's 4-bit subclasses (installed on the first
+compressed-tensors or fp8-to-NF4 load, then kept) and the Mistral FP8 wrapper took one argument and raised TypeError."""
+
+import inspect
+
+import pytest
+
+torch = pytest.importorskip("torch")
+pytest.importorskip("transformers")
+
+from transformers.quantizers.base import HfQuantizer
+
+if not hasattr(HfQuantizer, "update_weight_conversions"):
+ pytest.skip("transformers has no update_weight_conversions hook", allow_module_level = True)
+
+import unsloth # noqa: F401,E402
+from transformers import BitsAndBytesConfig, LlamaConfig, LlamaForCausalLM # noqa: E402
+from transformers.quantizers import AutoHfQuantizer # noqa: E402
+from transformers.quantizers import auto as quantizers_auto # noqa: E402
+
+from unsloth.models import compressed_tensors_bnb, fp8_to_nf4, mistral_format # noqa: E402
+
+PASSES_MODEL = "model" in inspect.signature(HfQuantizer.update_weight_conversions).parameters
+
+
+def _model():
+ from accelerate import init_empty_weights
+ with init_empty_weights():
+ return LlamaForCausalLM(
+ LlamaConfig(
+ hidden_size = 8,
+ num_hidden_layers = 1,
+ num_attention_heads = 2,
+ intermediate_size = 8,
+ vocab_size = 16,
+ )
+ )
+
+
+def _call(quantizer, model):
+ # The way the installed transformers calls it (conversion_mapping.get_model_conversion_mapping).
+ if PASSES_MODEL:
+ return quantizer.update_weight_conversions([], model)
+ return quantizer.update_weight_conversions([])
+
+
+def _accepts_extra_positional(fn):
+ return any(
+ p.kind is inspect.Parameter.VAR_POSITIONAL
+ for p in inspect.signature(fn).parameters.values()
+ )
+
+
+def _installed_bnb_overrides():
+ fp8_to_nf4.install_fp8_to_nf4_quantizer()
+ compressed_tensors_bnb.install_compressed_tensors_bnb_quantizer()
+ cls = quantizers_auto.AUTO_QUANTIZER_MAPPING["bitsandbytes_4bit"]
+ return [
+ k
+ for k in cls.__mro__
+ if k.__module__.startswith("unsloth.") and "update_weight_conversions" in vars(k)
+ ]
+
+
+def test_bnb_4bit_overrides_forward_extra_arguments():
+ overrides = _installed_bnb_overrides()
+ if not overrides:
+ pytest.skip("no Unsloth bitsandbytes 4-bit subclass is installed on this transformers")
+ for cls in overrides:
+ assert _accepts_extra_positional(vars(cls)["update_weight_conversions"]), cls.__qualname__
+
+
+def test_bnb_4bit_quantizer_takes_the_installed_call():
+ if not _installed_bnb_overrides():
+ pytest.skip("no Unsloth bitsandbytes 4-bit subclass is installed on this transformers")
+ quantizer = AutoHfQuantizer.from_config(
+ BitsAndBytesConfig(load_in_4bit = True), pre_quantized = False
+ )
+ assert isinstance(_call(quantizer, _model()), list)
+
+
+def test_mistral_fp8_dequantize_wrapper_takes_the_installed_call():
+ fp8 = pytest.importorskip("transformers.quantizers.quantizer_finegrained_fp8")
+ from transformers import FineGrainedFP8Config
+
+ if "update_weight_conversions" not in vars(fp8.FineGrainedFP8HfQuantizer):
+ pytest.skip("FineGrainedFP8HfQuantizer has no update_weight_conversions of its own")
+ with mistral_format._mistral_format_conversions():
+ wrapper = vars(fp8.FineGrainedFP8HfQuantizer)["update_weight_conversions"]
+ assert _accepts_extra_positional(wrapper)
+ quantizer = fp8.FineGrainedFP8HfQuantizer(FineGrainedFP8Config(dequantize = True))
+ quantizer.pre_quantized = True
+ assert isinstance(_call(quantizer, _model()), list)
diff --git a/unsloth/models/compressed_tensors_bnb.py b/unsloth/models/compressed_tensors_bnb.py
index c1785846bb4..f2cd6aa4088 100644
--- a/unsloth/models/compressed_tensors_bnb.py
+++ b/unsloth/models/compressed_tensors_bnb.py
@@ -1247,10 +1247,11 @@ def install_compressed_tensors_bnb_quantizer() -> bool:
for plan in self._unsloth_plan_dicts:
plan[key] = None
- def update_weight_conversions(self, weight_conversions):
+ def update_weight_conversions(self, weight_conversions, *args, **kwargs):
+ # transformers 5.21+ also passes the model.
ct_config = self._unsloth_ct_config
if ct_config is None:
- return super().update_weight_conversions(weight_conversions)
+ return super().update_weight_conversions(weight_conversions, *args, **kwargs)
if getattr(self, "_unsloth_keep_packed", False):
return (
list(weight_conversions)
@@ -1316,7 +1317,7 @@ def install_compressed_tensors_bnb_quantizer() -> bool:
parent = getattr(super(), "update_weight_conversions", None)
if parent is None:
return updated + list(self.get_weight_conversions())
- return parent(updated)
+ return parent(updated, *args, **kwargs)
mapping = getattr(quantizers_auto, "AUTO_QUANTIZER_MAPPING", None)
if mapping is None:
diff --git a/unsloth/models/fp8_to_nf4.py b/unsloth/models/fp8_to_nf4.py
index 11f36fa3c6a..0a9645c8ad2 100644
--- a/unsloth/models/fp8_to_nf4.py
+++ b/unsloth/models/fp8_to_nf4.py
@@ -1027,9 +1027,10 @@ def install_fp8_to_nf4_quantizer() -> bool:
_keep_fp8_embeddings(model, plan["modules_to_convert"])
return super()._process_model_before_weight_loading(model, **kwargs)
- def update_weight_conversions(self, weight_conversions):
+ def update_weight_conversions(self, weight_conversions, *args, **kwargs):
+ # transformers 5.21+ also passes the model.
if self._unsloth_fp8_plan is None or _STATE is None:
- return super().update_weight_conversions(weight_conversions)
+ return super().update_weight_conversions(weight_conversions, *args, **kwargs)
model = self._unsloth_fp8_model
self._unsloth_fp8_model = None
# Here, not before loading: only now are the caller's key_mapping renamings known.
@@ -1082,7 +1083,7 @@ def install_fp8_to_nf4_quantizer() -> bool:
)
)
self._unsloth_fp8_rebuilt = rebuilt_from
- return super().update_weight_conversions(updated)
+ return super().update_weight_conversions(updated, *args, **kwargs)
def _process_model_after_weight_loading(self, model, **kwargs):
global _STATE
diff --git a/unsloth/models/mistral_format.py b/unsloth/models/mistral_format.py
index 070892b4b2a..fbedeea17f9 100644
--- a/unsloth/models/mistral_format.py
+++ b/unsloth/models/mistral_format.py
@@ -628,12 +628,13 @@ def _mistral_format_conversions():
pass
if original_update is not None:
- def update_weight_conversions(self, weight_conversions):
+ def update_weight_conversions(self, weight_conversions, *args, **kwargs):
+ # transformers 5.21+ also passes the model.
if getattr(self.quantization_config, "dequantize", False):
weight_conversions = [
c for c in weight_conversions if not _is_expert_scale_merge(c)
]
- return original_update(self, weight_conversions)
+ return original_update(self, weight_conversions, *args, **kwargs)
fp8_quantizer.update_weight_conversions = update_weight_conversions
_active_conversions += 1