Intel support for Pixart-Sigma (#64)

* Fix IPEX attn_drop

* IPEX SDPA 4GB fix

Alchemist GPUs lack the 64 bit emulation needed for >4GB in this case. Attention slicing workaround copied from https://github.com/vladmandic/automatic/blob/master/modules/intel/ipex/attention.py

* Do 4GB fix only when necessary

~20% performance boost to not use the 4GB fix when not necessary

* Move attention.py inside utils

* Better 4GB estimation

* Remove unnecessary warning

PixArt/models/PixArt_blocks.py

@@ -15,18 +15,27 @@ import torch.nn.functional as F
 from timm.models.vision_transformer import Mlp, Attention as Attention_
 from einops import rearrange
 
+sdpa_32b = None
+Q_4GB_LIMIT = 32000000
+"""If q is greater than this, the operation will likely require >4GB VRAM, which will fail on Intel Arc Alchemist GPUs without a workaround."""
+# 2k   = 37 748 736
+# 1024 =  9 437 184
+# 2k model goes very slightly over 4GB
+
 from comfy import model_management
 if model_management.xformers_enabled():
     import xformers
     import xformers.ops
 else:
-    print("""
+    if model_management.xpu_available:
-########################################
+        import intel_extension_for_pytorch as ipex
- PixArt: Not using xformers!
+        import os
-  Expect images to be non-deterministic!
+        if not torch.xpu.has_fp64_dtype() and not os.environ.get('IPEX_FORCE_ATTENTION_SLICE', None):
-  Batch sizes > 1 are most likely broken
+            from ...utils.IPEX.attention import scaled_dot_product_attention_32_bit
-########################################
+            sdpa_32b = scaled_dot_product_attention_32_bit
-""")
+            print("Using IPEX 4GB SDPA workaround")
+        else:
+            print("No IPEX 4GB workaround")
 
 def modulate(x, shift, scale):
     return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
@@ -84,10 +93,17 @@ class MultiHeadCrossAttention(nn.Module):
                 for n in range(B - 1):
                     attn_mask = torch.block_diag(attn_mask, attn_mask_template)
 
-            x = torch.nn.functional.scaled_dot_product_attention(
+            p = getattr(self.attn_drop, "p", 0) # IPEX.optimize() will turn attn_drop into an Identity()
+
+            if sdpa_32b is not None and (q.element_size() * q.nelement()) > Q_4GB_LIMIT:
+                sdpa = sdpa_32b
+            else:
+                sdpa = torch.nn.functional.scaled_dot_product_attention
+
+            x = sdpa(
                 q, k, v,
                 attn_mask=attn_mask,
-                dropout_p=self.attn_drop.p
+                dropout_p=p
             ).permute(0, 2, 1, 3).contiguous()
         x = x.view(B, -1, C)
         x = self.proj(x)
@@ -193,9 +209,17 @@ class AttentionKVCompress(Attention_):
             )
         else:
             q, k, v = map(lambda t: t.transpose(1, 2),(q, k, v),)
-            x = torch.nn.functional.scaled_dot_product_attention(
+
+            p = getattr(self.attn_drop, "p", 0) # IPEX.optimize() will turn attn_drop into an Identity()
+            
+            if sdpa_32b is not None and (q.element_size() * q.nelement()) > Q_4GB_LIMIT:
+                sdpa = sdpa_32b
+            else:
+                sdpa = torch.nn.functional.scaled_dot_product_attention
+
+            x = sdpa(
                 q, k, v,
-                dropout_p=self.attn_drop.p,
+                dropout_p=p,
                 attn_mask=attn_bias
             ).transpose(1, 2).contiguous()
         x = x.view(B, N, C)

utils/IPEX/attention.py

@@ -0,0 +1,180 @@
+# Code lifted from https://github.com/Disty0/ipex_to_cuda/blob/main/attention.py
+# Thanks to Disty0! 
+
+import os
+import torch
+import intel_extension_for_pytorch as ipex # pylint: disable=import-error, unused-import
+from functools import cache
+
+# pylint: disable=protected-access, missing-function-docstring, line-too-long
+
+# ARC GPUs can't allocate more than 4GB to a single block so we slice the attetion layers
+
+sdpa_slice_trigger_rate = float(os.environ.get('IPEX_SDPA_SLICE_TRIGGER_RATE', 6))
+attention_slice_rate = float(os.environ.get('IPEX_ATTENTION_SLICE_RATE', 4))
+
+# Find something divisible with the input_tokens
+@cache
+def find_slice_size(slice_size, slice_block_size):
+    while (slice_size * slice_block_size) > attention_slice_rate:
+        slice_size = slice_size // 2
+        if slice_size <= 1:
+            slice_size = 1
+            break
+    return slice_size
+
+# Find slice sizes for SDPA
+@cache
+def find_sdpa_slice_sizes(query_shape, query_element_size):
+    if len(query_shape) == 3:
+        batch_size_attention, query_tokens, shape_three = query_shape
+        shape_four = 1
+    else:
+        batch_size_attention, query_tokens, shape_three, shape_four = query_shape
+
+    slice_block_size = query_tokens * shape_three * shape_four / 1024 / 1024 * query_element_size
+    block_size = batch_size_attention * slice_block_size
+
+    split_slice_size = batch_size_attention
+    split_2_slice_size = query_tokens
+    split_3_slice_size = shape_three
+
+    do_split = False
+    do_split_2 = False
+    do_split_3 = False
+
+    if block_size > sdpa_slice_trigger_rate:
+        do_split = True
+        split_slice_size = find_slice_size(split_slice_size, slice_block_size)
+        if split_slice_size * slice_block_size > attention_slice_rate:
+            slice_2_block_size = split_slice_size * shape_three * shape_four / 1024 / 1024 * query_element_size
+            do_split_2 = True
+            split_2_slice_size = find_slice_size(split_2_slice_size, slice_2_block_size)
+            if split_2_slice_size * slice_2_block_size > attention_slice_rate:
+                slice_3_block_size = split_slice_size * split_2_slice_size * shape_four / 1024 / 1024 * query_element_size
+                do_split_3 = True
+                split_3_slice_size = find_slice_size(split_3_slice_size, slice_3_block_size)
+
+    return do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size
+
+# Find slice sizes for BMM
+@cache
+def find_bmm_slice_sizes(input_shape, input_element_size, mat2_shape):
+    batch_size_attention, input_tokens, mat2_atten_shape = input_shape[0], input_shape[1], mat2_shape[2]
+    slice_block_size = input_tokens * mat2_atten_shape / 1024 / 1024 * input_element_size
+    block_size = batch_size_attention * slice_block_size
+
+    split_slice_size = batch_size_attention
+    split_2_slice_size = input_tokens
+    split_3_slice_size = mat2_atten_shape
+
+    do_split = False
+    do_split_2 = False
+    do_split_3 = False
+
+    if block_size > attention_slice_rate:
+        do_split = True
+        split_slice_size = find_slice_size(split_slice_size, slice_block_size)
+        if split_slice_size * slice_block_size > attention_slice_rate:
+            slice_2_block_size = split_slice_size * mat2_atten_shape / 1024 / 1024 * input_element_size
+            do_split_2 = True
+            split_2_slice_size = find_slice_size(split_2_slice_size, slice_2_block_size)
+            if split_2_slice_size * slice_2_block_size > attention_slice_rate:
+                slice_3_block_size = split_slice_size * split_2_slice_size / 1024 / 1024 * input_element_size
+                do_split_3 = True
+                split_3_slice_size = find_slice_size(split_3_slice_size, slice_3_block_size)
+
+    return do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size
+
+
+original_torch_bmm = torch.bmm
+def torch_bmm_32_bit(input, mat2, *, out=None):
+    if input.device.type != "xpu":
+        return original_torch_bmm(input, mat2, out=out)
+    do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size = find_bmm_slice_sizes(input.shape, input.element_size(), mat2.shape)
+
+    # Slice BMM
+    if do_split:
+        batch_size_attention, input_tokens, mat2_atten_shape = input.shape[0], input.shape[1], mat2.shape[2]
+        hidden_states = torch.zeros(input.shape[0], input.shape[1], mat2.shape[2], device=input.device, dtype=input.dtype)
+        for i in range(batch_size_attention // split_slice_size):
+            start_idx = i * split_slice_size
+            end_idx = (i + 1) * split_slice_size
+            if do_split_2:
+                for i2 in range(input_tokens // split_2_slice_size): # pylint: disable=invalid-name
+                    start_idx_2 = i2 * split_2_slice_size
+                    end_idx_2 = (i2 + 1) * split_2_slice_size
+                    if do_split_3:
+                        for i3 in range(mat2_atten_shape // split_3_slice_size): # pylint: disable=invalid-name
+                            start_idx_3 = i3 * split_3_slice_size
+                            end_idx_3 = (i3 + 1) * split_3_slice_size
+                            hidden_states[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] = original_torch_bmm(
+                                input[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
+                                mat2[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
+                                out=out
+                            )
+                    else:
+                        hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = original_torch_bmm(
+                            input[start_idx:end_idx, start_idx_2:end_idx_2],
+                            mat2[start_idx:end_idx, start_idx_2:end_idx_2],
+                            out=out
+                        )
+            else:
+                hidden_states[start_idx:end_idx] = original_torch_bmm(
+                    input[start_idx:end_idx],
+                    mat2[start_idx:end_idx],
+                    out=out
+                )
+        torch.xpu.synchronize(input.device)
+    else:
+        return original_torch_bmm(input, mat2, out=out)
+    return hidden_states
+
+original_scaled_dot_product_attention = torch.nn.functional.scaled_dot_product_attention
+def scaled_dot_product_attention_32_bit(query, key, value, attn_mask=None, dropout_p=0.0, is_causal=False, **kwargs):
+    if query.device.type != "xpu":
+        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs)
+    do_split, do_split_2, do_split_3, split_slice_size, split_2_slice_size, split_3_slice_size = find_sdpa_slice_sizes(query.shape, query.element_size())
+
+    # Slice SDPA
+    if do_split:
+        batch_size_attention, query_tokens, shape_three = query.shape[0], query.shape[1], query.shape[2]
+        hidden_states = torch.zeros(query.shape, device=query.device, dtype=query.dtype)
+        for i in range(batch_size_attention // split_slice_size):
+            start_idx = i * split_slice_size
+            end_idx = (i + 1) * split_slice_size
+            if do_split_2:
+                for i2 in range(query_tokens // split_2_slice_size): # pylint: disable=invalid-name
+                    start_idx_2 = i2 * split_2_slice_size
+                    end_idx_2 = (i2 + 1) * split_2_slice_size
+                    if do_split_3:
+                        for i3 in range(shape_three // split_3_slice_size): # pylint: disable=invalid-name
+                            start_idx_3 = i3 * split_3_slice_size
+                            end_idx_3 = (i3 + 1) * split_3_slice_size
+                            hidden_states[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] = original_scaled_dot_product_attention(
+                                query[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
+                                key[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
+                                value[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3],
+                                attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2, start_idx_3:end_idx_3] if attn_mask is not None else attn_mask,
+                                dropout_p=dropout_p, is_causal=is_causal, **kwargs
+                            )
+                    else:
+                        hidden_states[start_idx:end_idx, start_idx_2:end_idx_2] = original_scaled_dot_product_attention(
+                            query[start_idx:end_idx, start_idx_2:end_idx_2],
+                            key[start_idx:end_idx, start_idx_2:end_idx_2],
+                            value[start_idx:end_idx, start_idx_2:end_idx_2],
+                            attn_mask=attn_mask[start_idx:end_idx, start_idx_2:end_idx_2] if attn_mask is not None else attn_mask,
+                            dropout_p=dropout_p, is_causal=is_causal, **kwargs
+                        )
+            else:
+                hidden_states[start_idx:end_idx] = original_scaled_dot_product_attention(
+                    query[start_idx:end_idx],
+                    key[start_idx:end_idx],
+                    value[start_idx:end_idx],
+                    attn_mask=attn_mask[start_idx:end_idx] if attn_mask is not None else attn_mask,
+                    dropout_p=dropout_p, is_causal=is_causal, **kwargs
+                )
+        torch.xpu.synchronize(query.device)
+    else:
+        return original_scaled_dot_product_attention(query, key, value, attn_mask=attn_mask, dropout_p=dropout_p, is_causal=is_causal, **kwargs)
+    return hidden_states