Merge 45a2363e6a into 0904cc3fe5

This is an experimental WIP option that might not work in your workflow but
should lower memory usage if it does.

Currently only the VAE and the load image node will output in fp16 when
this option is turned on.

README.md

@@ -38,6 +38,8 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 
 ## Get Started
 
+### Local
+
 #### [Desktop Application](https://www.comfy.org/download)
 - The easiest way to get started.
 - Available on Windows & macOS.
@@ -49,8 +51,13 @@ ComfyUI lets you design and execute advanced stable diffusion pipelines using a
 #### [Manual Install](#manual-install-windows-linux)
 Supports all operating systems and GPU types (NVIDIA, AMD, Intel, Apple Silicon, Ascend).
 
-## [Examples](https://comfyanonymous.github.io/ComfyUI_examples/)
-See what ComfyUI can do with the [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
+### Cloud
+
+#### [Comfy Cloud](https://www.comfy.org/cloud)
+- Our official paid cloud version for those who can't afford local hardware.
+
+## Examples
+See what ComfyUI can do with the [newer template workflows](https://comfy.org/workflows) or old [example workflows](https://comfyanonymous.github.io/ComfyUI_examples/).
 
 ## Features
 - Nodes/graph/flowchart interface to experiment and create complex Stable Diffusion workflows without needing to code anything.

comfy/cli_args.py

@@ -83,6 +83,8 @@ fpte_group.add_argument("--fp16-text-enc", action="store_true", help="Store text
 fpte_group.add_argument("--fp32-text-enc", action="store_true", help="Store text encoder weights in fp32.")
 fpte_group.add_argument("--bf16-text-enc", action="store_true", help="Store text encoder weights in bf16.")
 
+parser.add_argument("--fp16-intermediates", action="store_true", help="Experimental: Use fp16 for intermediate tensors between nodes instead of fp32.")
+
 parser.add_argument("--force-channels-last", action="store_true", help="Force channels last format when inferencing the models.")
 
 parser.add_argument("--directml", type=int, nargs="?", metavar="DIRECTML_DEVICE", const=-1, help="Use torch-directml.")

comfy/float.py

@@ -55,6 +55,11 @@ def stochastic_rounding(value, dtype, seed=0):
     if dtype == torch.bfloat16:
         return value.to(dtype=torch.bfloat16)
     if dtype == torch.float8_e4m3fn or dtype == torch.float8_e5m2:
+        # MPS does not support FP8 dtypes — perform rounding on CPU and return the result there.
+        on_mps = value.device.type == "mps"
+        if on_mps:
+            value = value.cpu()
+
         generator = torch.Generator(device=value.device)
         generator.manual_seed(seed)
         output = torch.empty_like(value, dtype=dtype)
@@ -159,6 +164,12 @@ def stochastic_round_quantize_nvfp4(x, per_tensor_scale, pad_16x, seed=0):
         """Round up x to the nearest multiple."""
         return ((x + multiple - 1) // multiple) * multiple
 
+    # MPS does not support FP8 dtypes used for block scales — perform on CPU.
+    on_mps = x.device.type == "mps"
+    if on_mps:
+        x = x.cpu()
+        per_tensor_scale = per_tensor_scale.cpu() if isinstance(per_tensor_scale, torch.Tensor) else per_tensor_scale
+
     generator = torch.Generator(device=x.device)
     generator.manual_seed(seed)
 
@@ -179,6 +190,12 @@ def stochastic_round_quantize_nvfp4_by_block(x, per_tensor_scale, pad_16x, seed=
         """Round up x to the nearest multiple."""
         return ((x + multiple - 1) // multiple) * multiple
 
+    # MPS does not support FP8 dtypes used for block scales — perform on CPU.
+    on_mps = x.device.type == "mps"
+    if on_mps:
+        x = x.cpu()
+        per_tensor_scale = per_tensor_scale.cpu() if isinstance(per_tensor_scale, torch.Tensor) else per_tensor_scale
+
     orig_shape = x.shape
 
     # Handle padding
@@ -209,3 +226,39 @@ def stochastic_round_quantize_nvfp4_by_block(x, per_tensor_scale, pad_16x, seed=
         output_block[i:i + slice_size].copy_(block)
 
     return output_fp4, to_blocked(output_block, flatten=False)
+
+
+def stochastic_round_quantize_mxfp8_by_block(x, pad_32x, seed=0):
+    def roundup(x_val, multiple):
+        return ((x_val + multiple - 1) // multiple) * multiple
+
+    if pad_32x:
+        rows, cols = x.shape
+        padded_rows = roundup(rows, 32)
+        padded_cols = roundup(cols, 32)
+        if padded_rows != rows or padded_cols != cols:
+            x = torch.nn.functional.pad(x, (0, padded_cols - cols, 0, padded_rows - rows))
+
+    F8_E4M3_MAX = 448.0
+    E8M0_BIAS = 127
+    BLOCK_SIZE = 32
+
+    rows, cols = x.shape
+    x_blocked = x.reshape(rows, -1, BLOCK_SIZE)
+    max_abs = torch.amax(torch.abs(x_blocked), dim=-1)
+
+    # E8M0 block scales (power-of-2 exponents)
+    scale_needed = torch.clamp(max_abs.float() / F8_E4M3_MAX, min=2**(-127))
+    exp_biased = torch.clamp(torch.ceil(torch.log2(scale_needed)).to(torch.int32) + E8M0_BIAS, 0, 254)
+    block_scales_e8m0 = exp_biased.to(torch.uint8)
+
+    zero_mask = (max_abs == 0)
+    block_scales_f32 = (block_scales_e8m0.to(torch.int32) << 23).view(torch.float32)
+    block_scales_f32 = torch.where(zero_mask, torch.ones_like(block_scales_f32), block_scales_f32)
+
+    # Scale per-block then stochastic round
+    data_scaled = (x_blocked.float() / block_scales_f32.unsqueeze(-1)).reshape(rows, cols)
+    output_fp8 = stochastic_rounding(data_scaled, torch.float8_e4m3fn, seed=seed)
+
+    block_scales_e8m0 = torch.where(zero_mask, torch.zeros_like(block_scales_e8m0), block_scales_e8m0)
+    return output_fp8, to_blocked(block_scales_e8m0, flatten=False).view(torch.float8_e8m0fnu)

comfy/ldm/lightricks/vae/causal_video_autoencoder.py

@@ -11,6 +11,7 @@ from .causal_conv3d import CausalConv3d
 from .pixel_norm import PixelNorm
 from ..model import PixArtAlphaCombinedTimestepSizeEmbeddings
 import comfy.ops
+import comfy.model_management
 from comfy.ldm.modules.diffusionmodules.model import torch_cat_if_needed
 
 ops = comfy.ops.disable_weight_init
@@ -536,7 +537,7 @@ class Decoder(nn.Module):
                     mark_conv3d_ended(self.conv_out)
                 sample = self.conv_out(sample, causal=self.causal)
                 if sample is not None and sample.shape[2] > 0:
-                    output.append(sample)
+                    output.append(sample.to(comfy.model_management.intermediate_device()))
                 return
 
             up_block = self.up_blocks[idx]

comfy/model_management.py

@@ -1050,6 +1050,12 @@ def intermediate_device():
     else:
         return torch.device("cpu")
 
+def intermediate_dtype():
+    if args.fp16_intermediates:
+        return torch.float16
+    else:
+        return torch.float32
+
 def vae_device():
     if args.cpu_vae:
         return torch.device("cpu")
@@ -1712,6 +1718,19 @@ def supports_nvfp4_compute(device=None):
 
     return True
 
+def supports_mxfp8_compute(device=None):
+    if not is_nvidia():
+        return False
+
+    if torch_version_numeric < (2, 10):
+        return False
+
+    props = torch.cuda.get_device_properties(device)
+    if props.major < 10:
+        return False
+
+    return True
+
 def extended_fp16_support():
     # TODO: check why some models work with fp16 on newer torch versions but not on older
     if torch_version_numeric < (2, 7):

comfy/ops.py

@@ -857,6 +857,22 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
                             orig_shape=(self.out_features, self.in_features),
                         )
 
+                    elif self.quant_format == "mxfp8":
+                        # MXFP8: E8M0 block scales stored as uint8 in safetensors
+                        block_scale = self._load_scale_param(state_dict, prefix, "weight_scale", device, manually_loaded_keys,
+                                                             dtype=torch.uint8)
+
+                        if block_scale is None:
+                            raise ValueError(f"Missing MXFP8 block scales for layer {layer_name}")
+
+                        block_scale = block_scale.view(torch.float8_e8m0fnu)
+
+                        params = layout_cls.Params(
+                            scale=block_scale,
+                            orig_dtype=MixedPrecisionOps._compute_dtype,
+                            orig_shape=(self.out_features, self.in_features),
+                        )
+
                     elif self.quant_format == "nvfp4":
                         # NVFP4: tensor_scale (weight_scale_2) + block_scale (weight_scale)
                         tensor_scale = self._load_scale_param(state_dict, prefix, "weight_scale_2", device, manually_loaded_keys)
@@ -1006,12 +1022,15 @@ def mixed_precision_ops(quant_config={}, compute_dtype=torch.bfloat16, full_prec
 def pick_operations(weight_dtype, compute_dtype, load_device=None, disable_fast_fp8=False, fp8_optimizations=False, model_config=None):
     fp8_compute = comfy.model_management.supports_fp8_compute(load_device) # TODO: if we support more ops this needs to be more granular
     nvfp4_compute = comfy.model_management.supports_nvfp4_compute(load_device)
+    mxfp8_compute = comfy.model_management.supports_mxfp8_compute(load_device)
 
     if model_config and hasattr(model_config, 'quant_config') and model_config.quant_config:
         logging.info("Using mixed precision operations")
         disabled = set()
         if not nvfp4_compute:
             disabled.add("nvfp4")
+        if not mxfp8_compute:
+            disabled.add("mxfp8")
         if not fp8_compute:
             disabled.add("float8_e4m3fn")
             disabled.add("float8_e5m2")

comfy/quant_ops.py

@@ -43,6 +43,18 @@ except ImportError as e:
     def get_layout_class(name):
         return None
 
+_CK_MXFP8_AVAILABLE = False
+if _CK_AVAILABLE:
+    try:
+        from comfy_kitchen.tensor import TensorCoreMXFP8Layout as _CKMxfp8Layout
+        _CK_MXFP8_AVAILABLE = True
+    except ImportError:
+        logging.warning("comfy_kitchen does not support MXFP8, please update comfy_kitchen.")
+
+if not _CK_MXFP8_AVAILABLE:
+    class _CKMxfp8Layout:
+        pass
+
 import comfy.float
 
 # ==============================================================================
@@ -71,6 +83,12 @@ class _TensorCoreFP8LayoutBase(_CKFp8Layout):
         if not isinstance(scale, torch.Tensor):
             scale = torch.tensor(scale, device=tensor.device, dtype=torch.float32)
 
+        # MPS does not support FP8 dtypes — move to CPU for quantization.
+        on_mps = tensor.device.type == "mps"
+        if on_mps:
+            tensor = tensor.cpu()
+            scale = scale.cpu()
+
         if stochastic_rounding > 0:
             if inplace_ops:
                 tensor *= (1.0 / scale).to(tensor.dtype)
@@ -84,6 +102,31 @@ class _TensorCoreFP8LayoutBase(_CKFp8Layout):
         return qdata, params
 
 
+class TensorCoreMXFP8Layout(_CKMxfp8Layout):
+    @classmethod
+    def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
+        if tensor.dim() != 2:
+            raise ValueError(f"MXFP8 requires 2D tensor, got {tensor.dim()}D")
+
+        orig_dtype = tensor.dtype
+        orig_shape = tuple(tensor.shape)
+
+        padded_shape = cls.get_padded_shape(orig_shape)
+        needs_padding = padded_shape != orig_shape
+
+        if stochastic_rounding > 0:
+            qdata, block_scale = comfy.float.stochastic_round_quantize_mxfp8_by_block(tensor, pad_32x=needs_padding, seed=stochastic_rounding)
+        else:
+            qdata, block_scale = ck.quantize_mxfp8(tensor, pad_32x=needs_padding)
+
+        params = cls.Params(
+            scale=block_scale,
+            orig_dtype=orig_dtype,
+            orig_shape=orig_shape,
+        )
+        return qdata, params
+
+
 class TensorCoreNVFP4Layout(_CKNvfp4Layout):
     @classmethod
     def quantize(cls, tensor, scale=None, stochastic_rounding=0, inplace_ops=False):
@@ -137,6 +180,8 @@ register_layout_class("TensorCoreFP8Layout", TensorCoreFP8Layout)
 register_layout_class("TensorCoreFP8E4M3Layout", TensorCoreFP8E4M3Layout)
 register_layout_class("TensorCoreFP8E5M2Layout", TensorCoreFP8E5M2Layout)
 register_layout_class("TensorCoreNVFP4Layout", TensorCoreNVFP4Layout)
+if _CK_MXFP8_AVAILABLE:
+    register_layout_class("TensorCoreMXFP8Layout", TensorCoreMXFP8Layout)
 
 QUANT_ALGOS = {
     "float8_e4m3fn": {
@@ -157,6 +202,14 @@ QUANT_ALGOS = {
     },
 }
 
+if _CK_MXFP8_AVAILABLE:
+    QUANT_ALGOS["mxfp8"] = {
+        "storage_t": torch.float8_e4m3fn,
+        "parameters": {"weight_scale", "input_scale"},
+        "comfy_tensor_layout": "TensorCoreMXFP8Layout",
+        "group_size": 32,
+    }
+
 
 # ==============================================================================
 # Re-exports for backward compatibility

comfy/sd.py

@@ -871,13 +871,16 @@ class VAE:
                 pixels = torch.nn.functional.pad(pixels, (0, self.output_channels - pixels.shape[-1]), mode=mode, value=value)
         return pixels
 
+    def vae_output_dtype(self):
+        return model_management.intermediate_dtype()
+
     def decode_tiled_(self, samples, tile_x=64, tile_y=64, overlap = 16):
         steps = samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x, tile_y, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x // 2, tile_y * 2, overlap)
         steps += samples.shape[0] * comfy.utils.get_tiled_scale_steps(samples.shape[3], samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         output = self.process_output(
             (comfy.utils.tiled_scale(samples, decode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
             comfy.utils.tiled_scale(samples, decode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = self.upscale_ratio, output_device=self.output_device, pbar = pbar) +
@@ -887,16 +890,16 @@ class VAE:
 
     def decode_tiled_1d(self, samples, tile_x=256, overlap=32):
         if samples.ndim == 3:
-            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         else:
             og_shape = samples.shape
             samples = samples.reshape((og_shape[0], og_shape[1] * og_shape[2], -1))
-            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).float()
+            decode_fn = lambda a: self.first_stage_model.decode(a.reshape((-1, og_shape[1], og_shape[2], a.shape[-1])).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
 
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, output_device=self.output_device))
 
     def decode_tiled_3d(self, samples, tile_t=999, tile_x=32, tile_y=32, overlap=(1, 8, 8)):
-        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).float()
+        decode_fn = lambda a: self.first_stage_model.decode(a.to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return self.process_output(comfy.utils.tiled_scale_multidim(samples, decode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.upscale_ratio, out_channels=self.output_channels, index_formulas=self.upscale_index_formula, output_device=self.output_device))
 
     def encode_tiled_(self, pixel_samples, tile_x=512, tile_y=512, overlap = 64):
@@ -905,7 +908,7 @@ class VAE:
         steps += pixel_samples.shape[0] * comfy.utils.get_tiled_scale_steps(pixel_samples.shape[3], pixel_samples.shape[2], tile_x * 2, tile_y // 2, overlap)
         pbar = comfy.utils.ProgressBar(steps)
 
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         samples = comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x, tile_y, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x * 2, tile_y // 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
         samples += comfy.utils.tiled_scale(pixel_samples, encode_fn, tile_x // 2, tile_y * 2, overlap, upscale_amount = (1/self.downscale_ratio), out_channels=self.latent_channels, output_device=self.output_device, pbar=pbar)
@@ -914,7 +917,7 @@ class VAE:
 
     def encode_tiled_1d(self, samples, tile_x=256 * 2048, overlap=64 * 2048):
         if self.latent_dim == 1:
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
             out_channels = self.latent_channels
             upscale_amount = 1 / self.downscale_ratio
         else:
@@ -923,7 +926,7 @@ class VAE:
             tile_x = tile_x // extra_channel_size
             overlap = overlap // extra_channel_size
             upscale_amount = 1 / self.downscale_ratio
-            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).float()
+            encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).reshape(1, out_channels, -1).to(dtype=self.vae_output_dtype())
 
         out = comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_x,), overlap=overlap, upscale_amount=upscale_amount, out_channels=out_channels, output_device=self.output_device)
         if self.latent_dim == 1:
@@ -932,7 +935,7 @@ class VAE:
             return out.reshape(samples.shape[0], self.latent_channels, extra_channel_size, -1)
 
     def encode_tiled_3d(self, samples, tile_t=9999, tile_x=512, tile_y=512, overlap=(1, 64, 64)):
-        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).float()
+        encode_fn = lambda a: self.first_stage_model.encode((self.process_input(a)).to(self.vae_dtype).to(self.device)).to(dtype=self.vae_output_dtype())
         return comfy.utils.tiled_scale_multidim(samples, encode_fn, tile=(tile_t, tile_x, tile_y), overlap=overlap, upscale_amount=self.downscale_ratio, out_channels=self.latent_channels, downscale=True, index_formulas=self.downscale_index_formula, output_device=self.output_device)
 
     def decode(self, samples_in, vae_options={}):
@@ -950,9 +953,9 @@ class VAE:
 
             for x in range(0, samples_in.shape[0], batch_number):
                 samples = samples_in[x:x+batch_number].to(self.vae_dtype).to(self.device)
-                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).float())
+                out = self.process_output(self.first_stage_model.decode(samples, **vae_options).to(self.output_device).to(dtype=self.vae_output_dtype()))
                 if pixel_samples is None:
-                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    pixel_samples = torch.empty((samples_in.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 pixel_samples[x:x+batch_number] = out
         except Exception as e:
             model_management.raise_non_oom(e)
@@ -1025,9 +1028,9 @@ class VAE:
             samples = None
             for x in range(0, pixel_samples.shape[0], batch_number):
                 pixels_in = self.process_input(pixel_samples[x:x + batch_number]).to(self.vae_dtype).to(self.device)
-                out = self.first_stage_model.encode(pixels_in).to(self.output_device).float()
+                out = self.first_stage_model.encode(pixels_in).to(self.output_device).to(dtype=self.vae_output_dtype())
                 if samples is None:
-                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device)
+                    samples = torch.empty((pixel_samples.shape[0],) + tuple(out.shape[1:]), device=self.output_device, dtype=self.vae_output_dtype())
                 samples[x:x + batch_number] = out
 
         except Exception as e:

middleware/cache_middleware.py

@@ -32,7 +32,7 @@ async def cache_control(
     )
 
     if request.path.endswith(".js") or request.path.endswith(".css") or is_entry_point:
-        response.headers.setdefault("Cache-Control", "no-cache")
+        response.headers.setdefault("Cache-Control", "no-store")
         return response
 
     # Early return for non-image files - no cache headers needed

nodes.py

@@ -1724,6 +1724,8 @@ class LoadImage:
         output_masks = []
         w, h = None, None
 
+        dtype = comfy.model_management.intermediate_dtype()
+
         for i in ImageSequence.Iterator(img):
             i = node_helpers.pillow(ImageOps.exif_transpose, i)
 
@@ -1748,8 +1750,8 @@ class LoadImage:
                 mask = 1. - torch.from_numpy(mask)
             else:
                 mask = torch.zeros((64,64), dtype=torch.float32, device="cpu")
-            output_images.append(image)
-            output_masks.append(mask.unsqueeze(0))
+            output_images.append(image.to(dtype=dtype))
+            output_masks.append(mask.unsqueeze(0).to(dtype=dtype))
 
             if img.format == "MPO":
                 break  # ignore all frames except the first one for MPO format

requirements.txt

@@ -1,4 +1,4 @@
-comfyui-frontend-package==1.41.19
+comfyui-frontend-package==1.41.20
 comfyui-workflow-templates==0.9.21
 comfyui-embedded-docs==0.4.3
 torch
@@ -23,7 +23,7 @@ SQLAlchemy
 filelock
 av>=14.2.0
 comfy-kitchen>=0.2.8
-comfy-aimdo>=0.2.11
+comfy-aimdo>=0.2.12
 requests
 simpleeval>=1.0.0
 blake3

server.py

@@ -310,7 +310,7 @@ class PromptServer():
         @routes.get("/")
         async def get_root(request):
             response = web.FileResponse(os.path.join(self.web_root, "index.html"))
-            response.headers['Cache-Control'] = 'no-cache'
+            response.headers['Cache-Control'] = 'no-store, must-revalidate'
             response.headers["Pragma"] = "no-cache"
             response.headers["Expires"] = "0"
             return response

tests-unit/server_test/test_cache_control.py

@@ -28,31 +28,31 @@ CACHE_SCENARIOS = [
     },
     # JavaScript/CSS scenarios
     {
-        "name": "js_no_cache",
+        "name": "js_no_store",
         "path": "/script.js",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "css_no_cache",
+        "name": "css_no_store",
         "path": "/styles.css",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "index_json_no_cache",
+        "name": "index_json_no_store",
         "path": "/api/index.json",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     {
-        "name": "localized_index_json_no_cache",
+        "name": "localized_index_json_no_store",
         "path": "/templates/index.zh.json",
         "status": 200,
-        "expected_cache": "no-cache",
+        "expected_cache": "no-store",
         "should_have_header": True,
     },
     # Non-matching files

tests/test_fp8_mps.py

@@ -0,0 +1,147 @@
+"""
+Tests for FP8 quantization on MPS (Apple Silicon) devices.
+
+MPS does not natively support float8_e4m3fn or float8_e5m2 dtypes.
+These tests verify that:
+  1. FP8 operations correctly fall back to CPU when on MPS.
+  2. The round-trip (quantize on CPU -> result on original device) is numerically sound.
+  3. No "Placeholder storage has not been allocated on MPS device!" errors occur.
+"""
+import pytest
+import torch
+
+import comfy.float
+from comfy.quant_ops import TensorCoreFP8E4M3Layout, TensorCoreFP8E5M2Layout
+
+# Skip the entire module if MPS is not available
+pytestmark = pytest.mark.skipif(
+    not torch.backends.mps.is_available(),
+    reason="MPS backend not available"
+)
+
+# ── helpers ──────────────────────────────────────────────────────────────────
+
+def _make_mps_tensor(shape=(256, 256), dtype=torch.float32):
+    return torch.randn(shape, device="mps", dtype=dtype)
+
+
+# ── Tests for comfy.float ────────────────────────────────────────────────────
+
+class TestStochasticRoundingMPS:
+    """Tests for comfy.float.stochastic_rounding on MPS device."""
+
+    def test_stochastic_rounding_fp8_e4m3fn_on_mps(self):
+        """stochastic_rounding must not crash when input is on MPS and target dtype is float8_e4m3fn."""
+        x = _make_mps_tensor((64, 64), dtype=torch.float32)
+        result = comfy.float.stochastic_rounding(x, dtype=torch.float8_e4m3fn, seed=42)
+
+        assert result.dtype == torch.float8_e4m3fn
+        assert result.shape == x.shape
+
+    def test_stochastic_rounding_fp8_e5m2_on_mps(self):
+        """stochastic_rounding must not crash when input is on MPS and target dtype is float8_e5m2."""
+        x = _make_mps_tensor((64, 64), dtype=torch.float32)
+        result = comfy.float.stochastic_rounding(x, dtype=torch.float8_e5m2, seed=42)
+
+        assert result.dtype == torch.float8_e5m2
+        assert result.shape == x.shape
+
+    def test_stochastic_rounding_fp8_result_on_cpu(self):
+        """Result of FP8 rounding from MPS input should be on CPU (since MPS can't hold FP8)."""
+        x = _make_mps_tensor((32, 32), dtype=torch.float32)
+        result = comfy.float.stochastic_rounding(x, dtype=torch.float8_e4m3fn, seed=42)
+
+        # FP8 tensors cannot live on MPS, so result must be on CPU
+        assert result.device.type == "cpu"
+
+    def test_stochastic_rounding_non_fp8_still_works(self):
+        """Non-FP8 dtypes on MPS must still work as before (no regression)."""
+        x = _make_mps_tensor((32, 32), dtype=torch.float32)
+
+        r16 = comfy.float.stochastic_rounding(x, dtype=torch.float16, seed=0)
+        assert r16.dtype == torch.float16
+        assert r16.device.type == "mps"
+
+        rbf16 = comfy.float.stochastic_rounding(x, dtype=torch.bfloat16, seed=0)
+        assert rbf16.dtype == torch.bfloat16
+        assert rbf16.device.type == "mps"
+
+    def test_stochastic_rounding_fp8_numerical_sanity(self):
+        """FP8 round-trip (float32 -> fp8 -> float32) should have bounded error."""
+        x = torch.randn(128, 128, device="mps", dtype=torch.float32)
+        x_clamped = torch.clamp(x, min=-448, max=448)  # FP8 e4m3fn range
+
+        fp8 = comfy.float.stochastic_rounding(x_clamped, dtype=torch.float8_e4m3fn, seed=123)
+        # Convert back to float32 for comparison
+        reconstructed = fp8.to(torch.float32)
+
+        # Max relative error should be bounded (FP8 e4m3fn has ~0.125 relative precision)
+        x_cpu = x_clamped.cpu()
+        max_abs_err = (reconstructed - x_cpu).abs().max().item()
+        # FP8 e4m3fn max value is 448, min subnormal ~0.001953
+        # For random normal data, error should be well under 1.0
+        assert max_abs_err < 2.0, f"FP8 round-trip error too large: {max_abs_err}"
+
+
+class TestManualStochasticRoundMPS:
+    """Tests for comfy.float.manual_stochastic_round_to_float8 on MPS device."""
+
+    def test_manual_round_fp8_on_mps_tensor(self):
+        """stochastic_rounding handles MPS generator internally without 'Placeholder storage' error."""
+        x = _make_mps_tensor((16, 16), dtype=torch.float32)
+        result = comfy.float.stochastic_rounding(x, dtype=torch.float8_e4m3fn, seed=42)
+        assert result.dtype == torch.float8_e4m3fn
+
+
+class TestNVFP4StochasticRoundMPS:
+    """Tests for NVFP4 stochastic rounding on MPS - also creates FP8 tensors internally."""
+
+    def test_nvfp4_stochastic_round_on_mps(self):
+        """stochastic_round_quantize_nvfp4 creates FP8 block scales internally."""
+        # NVFP4 requires 2D input with dimensions divisible by 16
+        x = torch.randn(32, 32, device="mps", dtype=torch.float32)
+        scale = torch.tensor(1.0, device="mps", dtype=torch.float32)
+
+        # This should not crash - internally creates float8_e4m3fn block scales
+        qdata, block_scale = comfy.float.stochastic_round_quantize_nvfp4(
+            x, scale, pad_16x=False, seed=42
+        )
+        assert qdata.dtype == torch.uint8
+
+
+# ── Tests for comfy.quant_ops (integration) ──────────────────────────────────
+
+class TestQuantOpsMPS:
+    """Tests for the quantization ops layer that calls into comfy.float."""
+
+    def test_fp8_layout_quantize_on_mps(self):
+        """TensorCoreFP8E4M3Layout.quantize must work with MPS tensors."""
+        x = _make_mps_tensor((64, 64), dtype=torch.bfloat16)
+        qdata, params = TensorCoreFP8E4M3Layout.quantize(
+            x, scale="recalculate", stochastic_rounding=42
+        )
+
+        assert qdata.dtype == torch.float8_e4m3fn
+        assert params.orig_dtype == torch.bfloat16
+
+    def test_fp8_layout_quantize_without_stochastic_on_mps(self):
+        """TensorCoreFP8E4M3Layout.quantize with stochastic_rounding=0 uses ck.quantize_per_tensor_fp8."""
+        x = _make_mps_tensor((64, 64), dtype=torch.bfloat16)
+        qdata, params = TensorCoreFP8E4M3Layout.quantize(
+            x, scale="recalculate", stochastic_rounding=0
+        )
+
+        assert qdata.dtype == torch.float8_e4m3fn
+
+    def test_fp8_e5m2_layout_quantize_on_mps(self):
+        """TensorCoreFP8E5M2Layout.quantize must work with MPS tensors."""
+        x = _make_mps_tensor((64, 64), dtype=torch.float32)
+        qdata, params = TensorCoreFP8E5M2Layout.quantize(
+            x, scale="recalculate", stochastic_rounding=42
+        )
+
+        assert qdata.dtype == torch.float8_e5m2
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v", "--tb=short"])