add NV option for Random number generator source setting, which allows to generate same pictures on CPU/AMD/Mac as on NVidia videocards.

modules/devices.py

@@ -3,7 +3,7 @@ import contextlib
 from functools import lru_cache
 
 import torch
-from modules import errors
+from modules import errors, rng_philox
 
 if sys.platform == "darwin":
     from modules import mac_specific
@@ -90,23 +90,58 @@ def cond_cast_float(input):
     return input.float() if unet_needs_upcast else input
 
 
+nv_rng = None
+
+
 def randn(seed, shape):
     from modules.shared import opts
 
-    torch.manual_seed(seed)
+    manual_seed(seed)
+
+    if opts.randn_source == "NV":
+        return torch.asarray(nv_rng.randn(shape), device=device)
+
     if opts.randn_source == "CPU" or device.type == 'mps':
         return torch.randn(shape, device=cpu).to(device)
+
     return torch.randn(shape, device=device)
 
 
+def randn_like(x):
+    from modules.shared import opts
+
+    if opts.randn_source == "NV":
+        return torch.asarray(nv_rng.randn(x.shape), device=x.device, dtype=x.dtype)
+
+    if opts.randn_source == "CPU" or x.device.type == 'mps':
+        return torch.randn_like(x, device=cpu).to(x.device)
+
+    return torch.randn_like(x)
+
+
 def randn_without_seed(shape):
     from modules.shared import opts
 
+    if opts.randn_source == "NV":
+        return torch.asarray(nv_rng.randn(shape), device=device)
+
     if opts.randn_source == "CPU" or device.type == 'mps':
         return torch.randn(shape, device=cpu).to(device)
+
     return torch.randn(shape, device=device)
 
 
+def manual_seed(seed):
+    from modules.shared import opts
+
+    if opts.randn_source == "NV":
+        global nv_rng
+        nv_rng = rng_philox.Generator(seed)
+        return
+
+    torch.manual_seed(seed)
+
+
 def autocast(disable=False):
     from modules import shared
 

modules/processing.py

@@ -492,7 +492,7 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see
         noise_shape = shape if seed_resize_from_h <= 0 or seed_resize_from_w <= 0 else (shape[0], seed_resize_from_h//8, seed_resize_from_w//8)
 
         subnoise = None
-        if subseeds is not None:
+        if subseeds is not None and subseed_strength != 0:
             subseed = 0 if i >= len(subseeds) else subseeds[i]
 
             subnoise = devices.randn(subseed, noise_shape)
@@ -524,7 +524,7 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see
             cnt = p.sampler.number_of_needed_noises(p)
 
             if eta_noise_seed_delta > 0:
-                torch.manual_seed(seed + eta_noise_seed_delta)
+                devices.manual_seed(seed + eta_noise_seed_delta)
 
             for j in range(cnt):
                 sampler_noises[j].append(devices.randn_without_seed(tuple(noise_shape)))
@@ -636,7 +636,7 @@ def create_infotext(p, all_prompts, all_seeds, all_subseeds, comments=None, iter
         "Token merging ratio": None if token_merging_ratio == 0 else token_merging_ratio,
         "Token merging ratio hr": None if not enable_hr or token_merging_ratio_hr == 0 else token_merging_ratio_hr,
         "Init image hash": getattr(p, 'init_img_hash', None),
-        "RNG": opts.randn_source if opts.randn_source != "GPU" else None,
+        "RNG": opts.randn_source if opts.randn_source != "GPU" and opts.randn_source != "NV" else None,
         "NGMS": None if p.s_min_uncond == 0 else p.s_min_uncond,
         **p.extra_generation_params,
         "Version": program_version() if opts.add_version_to_infotext else None,

modules/rng_philox.py

@@ -0,0 +1,100 @@
+"""RNG imitiating torch cuda randn on CPU. You are welcome.
+
+Usage:
+
+```
+g = Generator(seed=0)
+print(g.randn(shape=(3, 4)))
+```
+
+Expected output:
+```
+[[-0.92466259 -0.42534415 -2.6438457   0.14518388]
+ [-0.12086647 -0.57972564 -0.62285122 -0.32838709]
+ [-1.07454231 -0.36314407 -1.67105067  2.26550497]]
+```
+"""
+
+import numpy as np
+
+philox_m = [0xD2511F53, 0xCD9E8D57]
+philox_w = [0x9E3779B9, 0xBB67AE85]
+
+two_pow32_inv = np.array([2.3283064e-10], dtype=np.float32)
+two_pow32_inv_2pi = np.array([2.3283064e-10 * 6.2831855], dtype=np.float32)
+
+
+def uint32(x):
+    """Converts (N,) np.uint64 array into (2, N) np.unit32 array."""
+    return np.moveaxis(x.view(np.uint32).reshape(-1, 2), 0, 1)
+
+
+def philox4_round(counter, key):
+    """A single round of the Philox 4x32 random number generator."""
+
+    v1 = uint32(counter[0].astype(np.uint64) * philox_m[0])
+    v2 = uint32(counter[2].astype(np.uint64) * philox_m[1])
+
+    counter[0] = v2[1] ^ counter[1] ^ key[0]
+    counter[1] = v2[0]
+    counter[2] = v1[1] ^ counter[3] ^ key[1]
+    counter[3] = v1[0]
+
+
+def philox4_32(counter, key, rounds=10):
+    """Generates 32-bit random numbers using the Philox 4x32 random number generator.
+
+    Parameters:
+        counter (numpy.ndarray): A 4xN array of 32-bit integers representing the counter values (offset into generation).
+        key (numpy.ndarray): A 2xN array of 32-bit integers representing the key values (seed).
+        rounds (int): The number of rounds to perform.
+
+    Returns:
+        numpy.ndarray: A 4xN array of 32-bit integers containing the generated random numbers.
+    """
+
+    for _ in range(rounds - 1):
+        philox4_round(counter, key)
+
+        key[0] = key[0] + philox_w[0]
+        key[1] = key[1] + philox_w[1]
+
+    philox4_round(counter, key)
+    return counter
+
+
+def box_muller(x, y):
+    """Returns just the first out of two numbers generated by Box–Muller transform algorithm."""
+    u = x.astype(np.float32) * two_pow32_inv + two_pow32_inv / 2
+    v = y.astype(np.float32) * two_pow32_inv_2pi + two_pow32_inv_2pi / 2
+
+    s = np.sqrt(-2.0 * np.log(u))
+
+    r1 = s * np.sin(v)
+    return r1.astype(np.float32)
+
+
+class Generator:
+    """RNG that produces same outputs as torch.randn(..., device='cuda') on CPU"""
+
+    def __init__(self, seed):
+        self.seed = seed
+        self.offset = 0
+
+    def randn(self, shape):
+        """Generate a sequence of n standard normal random variables using the Philox 4x32 random number generator and the Box-Muller transform."""
+
+        n = 1
+        for x in shape:
+            n *= x
+
+        counter = np.zeros((4, n), dtype=np.uint32)
+        counter[0] = self.offset
+        counter[2] = np.arange(n, dtype=np.uint32)  # up to 2^32 numbers can be generated - if you want more you'd need to spill into counter[3]
+        self.offset += 1
+
+        key = uint32(np.array([[self.seed] * n], dtype=np.uint64))
+
+        g = philox4_32(counter, key)
+
+        return box_muller(g[0], g[1]).reshape(shape)  # discard g[2] and g[3]

modules/sd_samplers_kdiffusion.py

@@ -260,10 +260,7 @@ class TorchHijack:
             if noise.shape == x.shape:
                 return noise
 
-        if opts.randn_source == "CPU" or x.device.type == 'mps':
-            return torch.randn_like(x, device=devices.cpu).to(x.device)
-        else:
-            return torch.randn_like(x)
+        return devices.randn_like(x)
 
 
 class KDiffusionSampler:

modules/shared.py

@@ -428,7 +428,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}).link("wiki", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#clip-skip").info("ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer"),
     "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),
     "auto_vae_precision": OptionInfo(True, "Automaticlly revert VAE to 32-bit floats").info("triggers when a tensor with NaNs is produced in VAE; disabling the option in this case will result in a black square image"),
-    "randn_source": OptionInfo("GPU", "Random number generator source.", gr.Radio, {"choices": ["GPU", "CPU"]}).info("changes seeds drastically; use CPU to produce the same picture across different videocard vendors"),
+    "randn_source": OptionInfo("GPU", "Random number generator source.", gr.Radio, {"choices": ["GPU", "CPU", "NV"]}).info("changes seeds drastically; use CPU to produce the same picture across different videocard vendors; use NV to produce same picture as on NVidia videocards"),
 }))
 
 options_templates.update(options_section(('sdxl', "Stable Diffusion XL"), {