From 153bc524bf9db76d723289f6420f418f63966972 Mon Sep 17 00:00:00 2001 From: ComfyUI Wiki Date: Sat, 10 Jan 2026 14:29:30 +0800 Subject: [PATCH 1/5] chore: update embedded docs to v0.4.0 (#11776) --- requirements.txt | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/requirements.txt b/requirements.txt index 7686a5f8a..6c1cd86d2 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,6 +1,6 @@ comfyui-frontend-package==1.36.13 comfyui-workflow-templates==0.7.69 -comfyui-embedded-docs==0.3.1 +comfyui-embedded-docs==0.4.0 torch torchsde torchvision From dc202a2e51bf7a6cd00e606b2d2941bc223f2ad2 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Fri, 9 Jan 2026 23:03:57 -0800 Subject: [PATCH 2/5] Properly save mixed ops. (#11772) --- comfy/ops.py | 26 ++++++++++++------- .../comfy_quant/test_mixed_precision.py | 6 ++--- 2 files changed, 20 insertions(+), 12 deletions(-) diff --git a/comfy/ops.py b/comfy/ops.py index 1cf22f0cc..9c0b54ff4 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -625,21 +625,29 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec missing_keys.remove(key) def state_dict(self, *args, destination=None, prefix="", **kwargs): - sd = super().state_dict(*args, destination=destination, prefix=prefix, **kwargs) - if isinstance(self.weight, QuantizedTensor): - layout_cls = self.weight._layout_cls + if destination is not None: + sd = destination + else: + sd = {} - # Check if it's any FP8 variant (E4M3 or E5M2) - if layout_cls in ("TensorCoreFP8E4M3Layout", "TensorCoreFP8E5M2Layout", "TensorCoreFP8Layout"): - sd["{}weight_scale".format(prefix)] = self.weight._params.scale - elif layout_cls == "TensorCoreNVFP4Layout": - sd["{}weight_scale_2".format(prefix)] = self.weight._params.scale - sd["{}weight_scale".format(prefix)] = self.weight._params.block_scale + if self.bias is not None: + sd["{}bias".format(prefix)] = self.bias + + if isinstance(self.weight, QuantizedTensor): + sd_out = self.weight.state_dict("{}weight".format(prefix)) + for k in sd_out: + sd[k] = sd_out[k] quant_conf = {"format": self.quant_format} if self._full_precision_mm_config: quant_conf["full_precision_matrix_mult"] = True sd["{}comfy_quant".format(prefix)] = torch.tensor(list(json.dumps(quant_conf).encode('utf-8')), dtype=torch.uint8) + + input_scale = getattr(self, 'input_scale', None) + if input_scale is not None: + sd["{}input_scale".format(prefix)] = input_scale + else: + sd["{}weight".format(prefix)] = self.weight return sd def _forward(self, input, weight, bias): diff --git a/tests-unit/comfy_quant/test_mixed_precision.py b/tests-unit/comfy_quant/test_mixed_precision.py index 7b2eac940..7c740491d 100644 --- a/tests-unit/comfy_quant/test_mixed_precision.py +++ b/tests-unit/comfy_quant/test_mixed_precision.py @@ -153,9 +153,9 @@ class TestMixedPrecisionOps(unittest.TestCase): state_dict2 = model.state_dict() # Verify layer1.weight is a QuantizedTensor with scale preserved - self.assertIsInstance(state_dict2["layer1.weight"], QuantizedTensor) - self.assertEqual(state_dict2["layer1.weight"]._params.scale.item(), 3.0) - self.assertEqual(state_dict2["layer1.weight"]._layout_cls, "TensorCoreFP8E4M3Layout") + self.assertTrue(torch.equal(state_dict2["layer1.weight"].view(torch.uint8), fp8_weight.view(torch.uint8))) + self.assertEqual(state_dict2["layer1.weight_scale"].item(), 3.0) + self.assertEqual(model.layer1.weight._layout_cls, "TensorCoreFP8E4M3Layout") # Verify non-quantized layers are standard tensors self.assertNotIsInstance(state_dict2["layer2.weight"], QuantizedTensor) From 6e4b1f9d00306fe14d7ca5adf2c7468d631b23d5 Mon Sep 17 00:00:00 2001 From: DELUXA Date: Sat, 10 Jan 2026 23:51:05 +0200 Subject: [PATCH 3/5] pythorch_attn_by_def_on_gfx1200 (#11793) --- comfy/model_management.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index e5de4a5b5..9d39be7b2 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -368,7 +368,7 @@ try: if any((a in arch) for a in ["gfx90a", "gfx942", "gfx1100", "gfx1101", "gfx1151"]): # TODO: more arches, TODO: gfx950 ENABLE_PYTORCH_ATTENTION = True if rocm_version >= (7, 0): - if any((a in arch) for a in ["gfx1201"]): + if any((a in arch) for a in ["gfx1200", "gfx1201"]): ENABLE_PYTORCH_ATTENTION = True if torch_version_numeric >= (2, 7) and rocm_version >= (6, 4): if any((a in arch) for a in ["gfx1200", "gfx1201", "gfx950"]): # TODO: more arches, "gfx942" gives error on pytorch nightly 2.10 1013 rocm7.0 From cd912963f17c9ae00ec12e1869293edb78720831 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 10 Jan 2026 14:31:31 -0800 Subject: [PATCH 4/5] Fix issue with t5 text encoder in fp4. (#11794) --- comfy/model_detection.py | 2 ++ comfy/sd.py | 4 ++-- 2 files changed, 4 insertions(+), 2 deletions(-) diff --git a/comfy/model_detection.py b/comfy/model_detection.py index 0853b3aec..aff5a50b9 100644 --- a/comfy/model_detection.py +++ b/comfy/model_detection.py @@ -237,6 +237,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None): else: dit_config["vec_in_dim"] = None + dit_config["num_heads"] = dit_config["hidden_size"] // sum(dit_config["axes_dim"]) + dit_config["depth"] = count_blocks(state_dict_keys, '{}double_blocks.'.format(key_prefix) + '{}.') dit_config["depth_single_blocks"] = count_blocks(state_dict_keys, '{}single_blocks.'.format(key_prefix) + '{}.') if '{}distilled_guidance_layer.0.norms.0.scale'.format(key_prefix) in state_dict_keys or '{}distilled_guidance_layer.norms.0.scale'.format(key_prefix) in state_dict_keys: #Chroma diff --git a/comfy/sd.py b/comfy/sd.py index 5a7221620..b689c0dfc 100644 --- a/comfy/sd.py +++ b/comfy/sd.py @@ -1059,9 +1059,9 @@ def detect_te_model(sd): return TEModel.JINA_CLIP_2 if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd: weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"] - if weight.shape[-1] == 4096: + if weight.shape[0] == 10240: return TEModel.T5_XXL - elif weight.shape[-1] == 2048: + elif weight.shape[0] == 5120: return TEModel.T5_XL if 'encoder.block.23.layer.1.DenseReluDense.wi.weight' in sd: return TEModel.T5_XXL_OLD From 2f642d5d9b48ad7cad13bbdd5f8adcf506f565a7 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Sat, 10 Jan 2026 14:40:42 -0800 Subject: [PATCH 5/5] Fix chroma fp8 te being treated as fp16. (#11795) --- comfy/text_encoders/cosmos.py | 2 +- comfy/text_encoders/genmo.py | 2 +- comfy/text_encoders/pixart_t5.py | 2 +- 3 files changed, 3 insertions(+), 3 deletions(-) diff --git a/comfy/text_encoders/cosmos.py b/comfy/text_encoders/cosmos.py index 448381fa9..f4b40ac68 100644 --- a/comfy/text_encoders/cosmos.py +++ b/comfy/text_encoders/cosmos.py @@ -36,7 +36,7 @@ def te(dtype_t5=None, t5_quantization_metadata=None): if t5_quantization_metadata is not None: model_options = model_options.copy() model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata - if dtype is None: + if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) return CosmosTEModel_ diff --git a/comfy/text_encoders/genmo.py b/comfy/text_encoders/genmo.py index 5daea8135..2d7a3fbce 100644 --- a/comfy/text_encoders/genmo.py +++ b/comfy/text_encoders/genmo.py @@ -32,7 +32,7 @@ def mochi_te(dtype_t5=None, t5_quantization_metadata=None): if t5_quantization_metadata is not None: model_options = model_options.copy() model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata - if dtype is None: + if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) return MochiTEModel_ diff --git a/comfy/text_encoders/pixart_t5.py b/comfy/text_encoders/pixart_t5.py index e5e5f18be..51c6e50c7 100644 --- a/comfy/text_encoders/pixart_t5.py +++ b/comfy/text_encoders/pixart_t5.py @@ -36,7 +36,7 @@ def pixart_te(dtype_t5=None, t5_quantization_metadata=None): if t5_quantization_metadata is not None: model_options = model_options.copy() model_options["t5xxl_quantization_metadata"] = t5_quantization_metadata - if dtype is None: + if dtype_t5 is not None: dtype = dtype_t5 super().__init__(device=device, dtype=dtype, model_options=model_options) return PixArtTEModel_