From e1d85e7577d8f6355bd4cb3449bcb0a7e5f80cb8 Mon Sep 17 00:00:00 2001 From: Qiacheng Li Date: Wed, 12 Nov 2025 12:21:05 -0800 Subject: [PATCH 1/5] Update README.md for Intel Arc GPU installation, remove IPEX (#10729) IPEX is no longer needed for Intel Arc GPUs. Removing instruction to setup ipex. --- README.md | 6 +----- 1 file changed, 1 insertion(+), 5 deletions(-) diff --git a/README.md b/README.md index 8142f595b..9e28803a2 100644 --- a/README.md +++ b/README.md @@ -242,7 +242,7 @@ RDNA 4 (RX 9000 series): ### Intel GPUs (Windows and Linux) -(Option 1) Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html) +Intel Arc GPU users can install native PyTorch with torch.xpu support using pip. More information can be found [here](https://pytorch.org/docs/main/notes/get_start_xpu.html) 1. To install PyTorch xpu, use the following command: @@ -252,10 +252,6 @@ This is the command to install the Pytorch xpu nightly which might have some per ```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/xpu``` -(Option 2) Alternatively, Intel GPUs supported by Intel Extension for PyTorch (IPEX) can leverage IPEX for improved performance. - -1. visit [Installation](https://intel.github.io/intel-extension-for-pytorch/index.html#installation?platform=gpu) for more information. - ### NVIDIA Nvidia users should install stable pytorch using this command: From 18e7d6dba5f1012d4cf09e8f777dc85d56ff25c0 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 13 Nov 2025 07:19:53 +1000 Subject: [PATCH 2/5] mm/mp: always unload re-used but modified models (#10724) The partial unloader path in model re-use flow skips straight to the actual unload without any check of the patching UUID. This means that if you do an upscale flow with a model patch on an existing model, it will not apply your patchings. Fix by delaying the partial_unload until after the uuid checks. This is done by making partial_unload a model of partial_load where extra_mem is -ve. --- comfy/model_management.py | 5 +---- comfy/model_patcher.py | 3 +++ 2 files changed, 4 insertions(+), 4 deletions(-) diff --git a/comfy/model_management.py b/comfy/model_management.py index d8913082a..a21df54b3 100644 --- a/comfy/model_management.py +++ b/comfy/model_management.py @@ -503,10 +503,7 @@ class LoadedModel: use_more_vram = lowvram_model_memory if use_more_vram == 0: use_more_vram = 1e32 - if use_more_vram > 0: - self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) - else: - self.model.partially_unload(self.model.offload_device, -use_more_vram, force_patch_weights=force_patch_weights) + self.model_use_more_vram(use_more_vram, force_patch_weights=force_patch_weights) real_model = self.model.model diff --git a/comfy/model_patcher.py b/comfy/model_patcher.py index 68b0a9192..cf1b0d441 100644 --- a/comfy/model_patcher.py +++ b/comfy/model_patcher.py @@ -928,6 +928,9 @@ class ModelPatcher: extra_memory += (used - self.model.model_loaded_weight_memory) self.patch_model(load_weights=False) + if extra_memory < 0 and not unpatch_weights: + self.partially_unload(self.offload_device, -extra_memory, force_patch_weights=force_patch_weights) + return 0 full_load = False if self.model.model_lowvram == False and self.model.model_loaded_weight_memory > 0: self.apply_hooks(self.forced_hooks, force_apply=True) From 1c7eaeca1013e4315f36e0d4d274faa106001121 Mon Sep 17 00:00:00 2001 From: rattus <46076784+rattus128@users.noreply.github.com> Date: Thu, 13 Nov 2025 07:20:53 +1000 Subject: [PATCH 3/5] qwen: reduce VRAM usage (#10725) Clean up a bunch of stacked and no-longer-needed tensors on the QWEN VRAM peak (currently FFN). With this I go from OOMing at B=37x1328x1328 to being able to succesfully run B=47 (RTX5090). --- comfy/ldm/qwen_image/model.py | 20 ++++++++++++-------- 1 file changed, 12 insertions(+), 8 deletions(-) diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.py index e5d0d17c1..427ea19c1 100644 --- a/comfy/ldm/qwen_image/model.py +++ b/comfy/ldm/qwen_image/model.py @@ -236,10 +236,10 @@ class QwenImageTransformerBlock(nn.Module): img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1) txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1) - img_normed = self.img_norm1(hidden_states) - img_modulated, img_gate1 = self._modulate(img_normed, img_mod1) - txt_normed = self.txt_norm1(encoder_hidden_states) - txt_modulated, txt_gate1 = self._modulate(txt_normed, txt_mod1) + img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1) + del img_mod1 + txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1) + del txt_mod1 img_attn_output, txt_attn_output = self.attn( hidden_states=img_modulated, @@ -248,16 +248,20 @@ class QwenImageTransformerBlock(nn.Module): image_rotary_emb=image_rotary_emb, transformer_options=transformer_options, ) + del img_modulated + del txt_modulated hidden_states = hidden_states + img_gate1 * img_attn_output encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output + del img_attn_output + del txt_attn_output + del img_gate1 + del txt_gate1 - img_normed2 = self.img_norm2(hidden_states) - img_modulated2, img_gate2 = self._modulate(img_normed2, img_mod2) + img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2) hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2)) - txt_normed2 = self.txt_norm2(encoder_hidden_states) - txt_modulated2, txt_gate2 = self._modulate(txt_normed2, txt_mod2) + txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2) encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2)) return encoder_hidden_states, hidden_states From 8b0b93df51d04f08eb779cb84dc331fa18b43ae8 Mon Sep 17 00:00:00 2001 From: comfyanonymous <121283862+comfyanonymous@users.noreply.github.com> Date: Wed, 12 Nov 2025 14:04:41 -0800 Subject: [PATCH 4/5] Update Python 3.14 compatibility notes in README (#10730) --- README.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/README.md b/README.md index 9e28803a2..f51807ad5 100644 --- a/README.md +++ b/README.md @@ -200,7 +200,7 @@ comfy install ## Manual Install (Windows, Linux) -Python 3.14 will work if you comment out the `kornia` dependency in the requirements.txt file (breaks the canny node) but it is not recommended. +Python 3.14 works but you may encounter issues with the torch compile node. The free threaded variant is still missing some dependencies. Python 3.13 is very well supported. If you have trouble with some custom node dependencies on 3.13 you can try 3.12 From 3b3ef9a77ac03ed516a45063f9736f33085cecca Mon Sep 17 00:00:00 2001 From: contentis Date: Thu, 13 Nov 2025 00:26:52 +0100 Subject: [PATCH 5/5] Quantized Ops fixes (#10715) * offload support, bug fixes, remove mixins * add readme --- QUANTIZATION.md | 168 +++++++++++++++++++++++++++++++++++++++++++++ comfy/ops.py | 37 ++++------ comfy/quant_ops.py | 39 ++++++++++- 3 files changed, 219 insertions(+), 25 deletions(-) create mode 100644 QUANTIZATION.md diff --git a/QUANTIZATION.md b/QUANTIZATION.md new file mode 100644 index 000000000..1693e13f3 --- /dev/null +++ b/QUANTIZATION.md @@ -0,0 +1,168 @@ +# The Comfy guide to Quantization + + +## How does quantization work? + +Quantization aims to map a high-precision value x_f to a lower precision format with minimal loss in accuracy. These smaller formats then serve to reduce the models memory footprint and increase throughput by using specialized hardware. + +When simply converting a value from FP16 to FP8 using the round-nearest method we might hit two issues: +- The dynamic range of FP16 (-65,504, 65,504) far exceeds FP8 formats like E4M3 (-448, 448) or E5M2 (-57,344, 57,344), potentially resulting in clipped values +- The original values are concentrated in a small range (e.g. -1,1) leaving many FP8-bits "unused" + +By using a scaling factor, we aim to map these values into the quantized-dtype range, making use of the full spectrum. One of the easiest approaches, and common, is using per-tensor absolute-maximum scaling. + +``` +absmax = max(abs(tensor)) +scale = amax / max_dynamic_range_low_precision + +# Quantization +tensor_q = (tensor / scale).to(low_precision_dtype) + +# De-Quantization +tensor_dq = tensor_q.to(fp16) * scale + +tensor_dq ~ tensor +``` + +Given that additional information (scaling factor) is needed to "interpret" the quantized values, we describe those as derived datatypes. + + +## Quantization in Comfy + +``` +QuantizedTensor (torch.Tensor subclass) + ↓ __torch_dispatch__ +Two-Level Registry (generic + layout handlers) + ↓ +MixedPrecisionOps + Metadata Detection +``` + +### Representation + +To represent these derived datatypes, ComfyUI uses a subclass of torch.Tensor to implements these using the `QuantizedTensor` class found in `comfy/quant_ops.py` + +A `Layout` class defines how a specific quantization format behaves: +- Required parameters +- Quantize method +- De-Quantize method + +```python +from comfy.quant_ops import QuantizedLayout + +class MyLayout(QuantizedLayout): + @classmethod + def quantize(cls, tensor, **kwargs): + # Convert to quantized format + qdata = ... + params = {'scale': ..., 'orig_dtype': tensor.dtype} + return qdata, params + + @staticmethod + def dequantize(qdata, scale, orig_dtype, **kwargs): + return qdata.to(orig_dtype) * scale +``` + +To then run operations using these QuantizedTensors we use two registry systems to define supported operations. +The first is a **generic registry** that handles operations common to all quantized formats (e.g., `.to()`, `.clone()`, `.reshape()`). + +The second registry is layout-specific and allows to implement fast-paths like nn.Linear. +```python +from comfy.quant_ops import register_layout_op + +@register_layout_op(torch.ops.aten.linear.default, MyLayout) +def my_linear(func, args, kwargs): + # Extract tensors, call optimized kernel + ... +``` +When `torch.nn.functional.linear()` is called with QuantizedTensor arguments, `__torch_dispatch__` automatically routes to the registered implementation. +For any unsupported operation, QuantizedTensor will fallback to call `dequantize` and dispatch using the high-precision implementation. + + +### Mixed Precision + +The `MixedPrecisionOps` class (lines 542-648 in `comfy/ops.py`) enables per-layer quantization decisions, allowing different layers in a model to use different precisions. This is activated when a model config contains a `layer_quant_config` dictionary that specifies which layers should be quantized and how. + +**Architecture:** + +```python +class MixedPrecisionOps(disable_weight_init): + _layer_quant_config = {} # Maps layer names to quantization configs + _compute_dtype = torch.bfloat16 # Default compute / dequantize precision +``` + +**Key mechanism:** + +The custom `Linear._load_from_state_dict()` method inspects each layer during model loading: +- If the layer name is **not** in `_layer_quant_config`: load weight as regular tensor in `_compute_dtype` +- If the layer name **is** in `_layer_quant_config`: + - Load weight as `QuantizedTensor` with the specified layout (e.g., `TensorCoreFP8Layout`) + - Load associated quantization parameters (scales, block_size, etc.) + +**Why it's needed:** + +Not all layers tolerate quantization equally. Sensitive operations like final projections can be kept in higher precision, while compute-heavy matmuls are quantized. This provides most of the performance benefits while maintaining quality. + +The system is selected in `pick_operations()` when `model_config.layer_quant_config` is present, making it the highest-priority operation mode. + + +## Checkpoint Format + +Quantized checkpoints are stored as standard safetensors files with quantized weight tensors and associated scaling parameters, plus a `_quantization_metadata` JSON entry describing the quantization scheme. + +The quantized checkpoint will contain the same layers as the original checkpoint but: +- The weights are stored as quantized values, sometimes using a different storage datatype. E.g. uint8 container for fp8. +- For each quantized weight a number of additional scaling parameters are stored alongside depending on the recipe. +- We store a metadata.json in the metadata of the final safetensor containing the `_quantization_metadata` describing which layers are quantized and what layout has been used. + +### Scaling Parameters details +We define 4 possible scaling parameters that should cover most recipes in the near-future: +- **weight_scale**: quantization scalers for the weights +- **weight_scale_2**: global scalers in the context of double scaling +- **pre_quant_scale**: scalers used for smoothing salient weights +- **input_scale**: quantization scalers for the activations + +| Format | Storage dtype | weight_scale | weight_scale_2 | pre_quant_scale | input_scale | +|--------|---------------|--------------|----------------|-----------------|-------------| +| float8_e4m3fn | float32 | float32 (scalar) | - | - | float32 (scalar) | + +You can find the defined formats in `comfy/quant_ops.py` (QUANT_ALGOS). + +### Quantization Metadata + +The metadata stored alongside the checkpoint contains: +- **format_version**: String to define a version of the standard +- **layers**: A dictionary mapping layer names to their quantization format. The format string maps to the definitions found in `QUANT_ALGOS`. + +Example: +```json +{ + "_quantization_metadata": { + "format_version": "1.0", + "layers": { + "model.layers.0.mlp.up_proj": "float8_e4m3fn", + "model.layers.0.mlp.down_proj": "float8_e4m3fn", + "model.layers.1.mlp.up_proj": "float8_e4m3fn" + } + } +} +``` + + +## Creating Quantized Checkpoints + +To create compatible checkpoints, use any quantization tool provided the output follows the checkpoint format described above and uses a layout defined in `QUANT_ALGOS`. + +### Weight Quantization + +Weight quantization is straightforward - compute the scaling factor directly from the weight tensor using the absolute maximum method described earlier. Each layer's weights are quantized independently and stored with their corresponding `weight_scale` parameter. + +### Calibration (for Activation Quantization) + +Activation quantization (e.g., for FP8 Tensor Core operations) requires `input_scale` parameters that cannot be determined from static weights alone. Since activation values depend on actual inputs, we use **post-training calibration (PTQ)**: + +1. **Collect statistics**: Run inference on N representative samples +2. **Track activations**: Record the absolute maximum (`amax`) of inputs to each quantized layer +3. **Compute scales**: Derive `input_scale` from collected statistics +4. **Store in checkpoint**: Save `input_scale` parameters alongside weights + +The calibration dataset should be representative of your target use case. For diffusion models, this typically means a diverse set of prompts and generation parameters. \ No newline at end of file diff --git a/comfy/ops.py b/comfy/ops.py index 96dffa85d..2a90a5ba2 100644 --- a/comfy/ops.py +++ b/comfy/ops.py @@ -77,7 +77,10 @@ def cast_bias_weight(s, input=None, dtype=None, device=None, bias_dtype=None, of # will add async-offload support to your cast and improve performance. if input is not None: if dtype is None: - dtype = input.dtype + if isinstance(input, QuantizedTensor): + dtype = input._layout_params["orig_dtype"] + else: + dtype = input.dtype if bias_dtype is None: bias_dtype = dtype if device is None: @@ -534,18 +537,7 @@ if CUBLAS_IS_AVAILABLE: # ============================================================================== # Mixed Precision Operations # ============================================================================== -from .quant_ops import QuantizedTensor - -QUANT_FORMAT_MIXINS = { - "float8_e4m3fn": { - "dtype": torch.float8_e4m3fn, - "layout_type": "TensorCoreFP8Layout", - "parameters": { - "weight_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), - "input_scale": torch.nn.Parameter(torch.zeros((), dtype=torch.float32), requires_grad=False), - } - } -} +from .quant_ops import QuantizedTensor, QUANT_ALGOS class MixedPrecisionOps(disable_weight_init): _layer_quant_config = {} @@ -596,23 +588,24 @@ class MixedPrecisionOps(disable_weight_init): if quant_format is None: raise ValueError(f"Unknown quantization format for layer {layer_name}") - mixin = QUANT_FORMAT_MIXINS[quant_format] - self.layout_type = mixin["layout_type"] + qconfig = QUANT_ALGOS[quant_format] + self.layout_type = qconfig["comfy_tensor_layout"] - scale_key = f"{prefix}weight_scale" + weight_scale_key = f"{prefix}weight_scale" layout_params = { - 'scale': state_dict.pop(scale_key, None), - 'orig_dtype': MixedPrecisionOps._compute_dtype + 'scale': state_dict.pop(weight_scale_key, None), + 'orig_dtype': MixedPrecisionOps._compute_dtype, + 'block_size': qconfig.get("group_size", None), } if layout_params['scale'] is not None: - manually_loaded_keys.append(scale_key) + manually_loaded_keys.append(weight_scale_key) self.weight = torch.nn.Parameter( - QuantizedTensor(weight.to(device=device, dtype=mixin["dtype"]), self.layout_type, layout_params), + QuantizedTensor(weight.to(device=device), self.layout_type, layout_params), requires_grad=False ) - for param_name, param_value in mixin["parameters"].items(): + for param_name in qconfig["parameters"]: param_key = f"{prefix}{param_name}" _v = state_dict.pop(param_key, None) if _v is None: @@ -643,7 +636,7 @@ class MixedPrecisionOps(disable_weight_init): if (getattr(self, 'layout_type', None) is not None and getattr(self, 'input_scale', None) is not None and not isinstance(input, QuantizedTensor)): - input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, fp8_dtype=self.weight.dtype) + input = QuantizedTensor.from_float(input, self.layout_type, scale=self.input_scale, dtype=self.weight.dtype) return self._forward(input, self.weight, self.bias) diff --git a/comfy/quant_ops.py b/comfy/quant_ops.py index c56e32a73..1d058bece 100644 --- a/comfy/quant_ops.py +++ b/comfy/quant_ops.py @@ -74,6 +74,12 @@ def _copy_layout_params(params): new_params[k] = v return new_params +def _copy_layout_params_inplace(src, dst, non_blocking=False): + for k, v in src.items(): + if isinstance(v, torch.Tensor): + dst[k].copy_(v, non_blocking=non_blocking) + else: + dst[k] = v class QuantizedLayout: """ @@ -318,13 +324,13 @@ def generic_to_dtype_layout(func, args, kwargs): def generic_copy_(func, args, kwargs): qt_dest = args[0] src = args[1] - + non_blocking = args[2] if len(args) > 2 else False if isinstance(qt_dest, QuantizedTensor): if isinstance(src, QuantizedTensor): # Copy from another quantized tensor - qt_dest._qdata.copy_(src._qdata) + qt_dest._qdata.copy_(src._qdata, non_blocking=non_blocking) qt_dest._layout_type = src._layout_type - qt_dest._layout_params = _copy_layout_params(src._layout_params) + _copy_layout_params_inplace(src._layout_params, qt_dest._layout_params, non_blocking=non_blocking) else: # Copy from regular tensor - just copy raw data qt_dest._qdata.copy_(src) @@ -336,6 +342,26 @@ def generic_copy_(func, args, kwargs): def generic_has_compatible_shallow_copy_type(func, args, kwargs): return True + +@register_generic_util(torch.ops.aten.empty_like.default) +def generic_empty_like(func, args, kwargs): + """Empty_like operation - creates an empty tensor with the same quantized structure.""" + qt = args[0] + if isinstance(qt, QuantizedTensor): + # Create empty tensor with same shape and dtype as the quantized data + hp_dtype = kwargs.pop('dtype', qt._layout_params["orig_dtype"]) + new_qdata = torch.empty_like(qt._qdata, **kwargs) + + # Handle device transfer for layout params + target_device = kwargs.get('device', new_qdata.device) + new_params = _move_layout_params_to_device(qt._layout_params, target_device) + + # Update orig_dtype if dtype is specified + new_params['orig_dtype'] = hp_dtype + + return QuantizedTensor(new_qdata, qt._layout_type, new_params) + return func(*args, **kwargs) + # ============================================================================== # FP8 Layout + Operation Handlers # ============================================================================== @@ -378,6 +404,13 @@ class TensorCoreFP8Layout(QuantizedLayout): def get_plain_tensors(cls, qtensor): return qtensor._qdata, qtensor._layout_params['scale'] +QUANT_ALGOS = { + "float8_e4m3fn": { + "storage_t": torch.float8_e4m3fn, + "parameters": {"weight_scale", "input_scale"}, + "comfy_tensor_layout": "TensorCoreFP8Layout", + }, +} LAYOUTS = { "TensorCoreFP8Layout": TensorCoreFP8Layout,