Support SegmentID when doing data prallel SPMD (#8425)

Author: JackCaoG

test/test_pallas_spmd.py

@@ -3,6 +3,7 @@
 import unittest
 
 import torch
+import numpy as np
 from torch import nn as nn
 
 import torch_xla
@@ -22,8 +23,24 @@
 
 class PallasTest(unittest.TestCase):
 
-  def _attention(self, q, k, v):
+  # This is to create a diagonal mask where only elements within the same segment
+  # can attend to each other. Since the mask is to mask out the unrelevant parts,
+  # therefore we use != instead of ==.
+  def _make_attention_mask_from_segment_ids(self, q_segment_ids,
+                                            kv_segment_ids):
+    return q_segment_ids.view(q_segment_ids.shape[0], 1,
+                              q_segment_ids.shape[1], 1) != kv_segment_ids.view(
+                                  kv_segment_ids.shape[0], 1, 1,
+                                  kv_segment_ids.shape[1])
+
+  def _attention(self, q, k, v, *, attn_mask=None, ab=None):
     attn_weight = q @ k.transpose(-2, -1)
+    if attn_mask is not None:
+      # Masked out the unrelevant parts.
+      attn_weight = attn_weight.masked_fill(attn_mask,
+                                            torch.finfo(attn_weight.dtype).min)
+    if ab is not None:
+      attn_weight = attn_weight + ab
     attn_weight = nn.functional.softmax(attn_weight, dim=-1)
     attn_output = attn_weight @ v
     return attn_output
@@ -98,6 +115,117 @@ def test_flash_attention_backward_spmd_data_parallel(self):
       self.assertTrue(torch.allclose(i[0].grad.cpu(), i[1].cpu(), atol=1e-05))
     jax.config.update('jax_default_matmul_precision', "default")
 
+  @unittest.skipIf(xr.device_type() != 'TPU' or tpu.version() < 3,
+                   "This test only works on TPUv3+.")
+  def test_flash_attention_wrapper_segment_ids_spmd(self):
+    from torch_xla.experimental.custom_kernel import flash_attention
+    from jax.experimental.pallas.ops.tpu.flash_attention import flash_attention as jax_flash_attention, SegmentIds
+    xs.set_global_mesh(xs.get_1d_mesh("data"))
+
+    q = torch.randn(3, 2, 128, 4)
+    k = torch.randn(3, 2, 128, 4)
+    v = torch.randn(3, 2, 128, 4)
+    zeros = torch.zeros(3, 32)
+    segment_ids = torch.cat([zeros, zeros + 1, zeros + 2, zeros + 3], dim=1)
+    segment_ids_xla = segment_ids.to("xla")
+    # only shard data dimension
+    o = flash_attention(
+        q.to("xla"),
+        k.to("xla"),
+        v.to("xla"),
+        False,
+        segment_ids_xla,
+        segment_ids.to("xla"),
+        partition_spec=("data", None, None, None))
+    self.assertEqual(
+        torch_xla._XLAC._get_xla_sharding_spec(o),
+        f"{{devices=[{xr.global_runtime_device_count()},1,1,1]0,1,2,3}}")
+
+    jax_q = jnp.array(q.numpy(), dtype=jnp.float32)
+    jax_k = jnp.array(k.numpy(), dtype=jnp.float32)
+    jax_v = jnp.array(v.numpy(), dtype=jnp.float32)
+    jax_segment_ids = jnp.array(segment_ids.numpy(), dtype=jnp.float32)
+    expected_o = torch.from_numpy(
+        np.array(
+            jax_flash_attention(
+                jax_q,
+                jax_k,
+                jax_v,
+                segment_ids=SegmentIds(jax_segment_ids, jax_segment_ids),
+            )))
+
+    self.assertTrue(torch.allclose(o.cpu(), expected_o.cpu(), atol=1e-05))
+    jax.config.update('jax_default_matmul_precision', "default")
+
+  @unittest.skipIf(xr.device_type() != 'TPU' or tpu.version() < 3,
+                   "This test only works on TPUv3+.")
+  def test_flash_attention_backward_segment_ids_spmd(self):
+    jax.config.update("jax_default_matmul_precision", "highest")
+    from torch_xla.experimental.custom_kernel import flash_attention
+    n_devices = xr.global_runtime_device_count()
+    xs.set_global_mesh(xs.get_1d_mesh("data"))
+
+    torch.manual_seed(42)
+    q = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
+    k = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
+    v = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
+    zeros = torch.zeros(4, 32).to("xla")
+    segment_ids = torch.cat([zeros, zeros + 1, zeros + 2, zeros + 3], dim=1)
+    q.retain_grad()
+    k.retain_grad()
+    v.retain_grad()
+
+    o = flash_attention(
+        q,
+        k,
+        v,
+        False,
+        segment_ids,
+        segment_ids,
+        partition_spec=("data", None, None, None))
+    loss = o.sum()
+    loss.backward()
+    q_grad = q.grad
+    k_grad = k.grad
+    v_grad = v.grad
+    self.assertEqual(
+        torch_xla._XLAC._get_xla_sharding_spec(o),
+        f"{{devices=[{n_devices},1,1,1]0,1,2,3}}")
+    self.assertEqual(
+        torch_xla._XLAC._get_xla_sharding_spec(q_grad),
+        f"{{devices=[{n_devices},1,1,1]0,1,2,3}}")
+    self.assertEqual(
+        torch_xla._XLAC._get_xla_sharding_spec(k_grad),
+        f"{{devices=[{n_devices},1,1,1]0,1,2,3}}")
+    self.assertEqual(
+        torch_xla._XLAC._get_xla_sharding_spec(v_grad),
+        f"{{devices=[{n_devices},1,1,1]0,1,2,3}}")
+    torch_xla.sync()
+
+    torch.manual_seed(42)
+    q = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
+    k = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
+    v = torch.randn(4, 2, 128, 8, requires_grad=True).to("xla")
+    zeros = torch.zeros(4, 32).to("xla")
+    segment_ids = torch.cat([zeros, zeros + 1, zeros + 2, zeros + 3], dim=1)
+    q.retain_grad()
+    k.retain_grad()
+    v.retain_grad()
+
+    o = self._attention(
+        q,
+        k,
+        v,
+        attn_mask=self._make_attention_mask_from_segment_ids(
+            segment_ids, segment_ids))
+    loss = o.sum()
+    loss.backward()
+    xm.mark_step()
+
+    for i in [(q, q_grad), (k, k_grad), (v, v_grad)]:
+      self.assertTrue(torch.allclose(i[0].grad.cpu(), i[1].cpu(), atol=1e-05))
+    jax.config.update("jax_default_matmul_precision", "default")
+
 
 if __name__ == '__main__':
   logging.getLogger().setLevel(logging.INFO)

torch_xla/experimental/custom_kernel.py

@@ -266,7 +266,15 @@ def forward(ctx, q, k, v, causal, q_segment_ids, kv_segment_ids, sm_scale, ab,
         dtypes.append(torch.float32)
 
     with torch.no_grad():
-      segment_ids, q_segment_ids, kv_segment_ids = FlashAttention.prepare_segment_ids(
+      if partition_spec is not None and q_segment_ids is not None and kv_segment_ids is not None:
+        # partition_spec is for q,k,v with shape [batch, num_head, seq_len, head_dim], segment id
+        # is of shape [batch, seq_len], hence we need to tweak it a bit
+        segment_id_partition_spec = (partition_spec[0], partition_spec[2])
+        q_segment_ids = xs.enable_manual_sharding(
+            q_segment_ids, segment_id_partition_spec, mesh=mesh).global_tensor
+        kv_segment_ids = xs.enable_manual_sharding(
+            kv_segment_ids, segment_id_partition_spec, mesh=mesh).global_tensor
+      segment_ids, q_segment_ids_fa, kv_segment_ids_fa = FlashAttention.prepare_segment_ids(
           q_segment_ids, kv_segment_ids)
       ctx.segment_ids = segment_ids
 
@@ -297,7 +305,7 @@ def forward(ctx, q, k, v, causal, q_segment_ids, kv_segment_ids, sm_scale, ab,
       if ab is not None:
         args += [ab]
       if segment_ids is not None:
-        args += [q_segment_ids, kv_segment_ids]
+        args += [q_segment_ids_fa, kv_segment_ids_fa]
       o = torch_xla._XLAC._xla_tpu_custom_call(args, payload, shapes, dtypes)
 
       if not save_residuals:
@@ -319,20 +327,23 @@ def forward(ctx, q, k, v, causal, q_segment_ids, kv_segment_ids, sm_scale, ab,
       m = xs.disable_manual_sharding(
           m, partition_spec[0:3], ctx.full_shape[0:3], mesh=mesh).global_tensor
 
-    ctx.save_for_backward(full_q, full_k, full_v, o, l, m, q_segment_ids,
-                          kv_segment_ids, full_ab)
+    # q_segment_ids and kv_segment_ids are sharded here if partition_spec is provided
+    # but it should be OK as the backward will use the same partition_spec
+    ctx.save_for_backward(full_q, full_k, full_v, o, l, m, q_segment_ids_fa,
+                          kv_segment_ids_fa, full_ab)
     return o
 
   @staticmethod
   def backward(ctx, grad_output):
     from jax.experimental.pallas.ops.tpu.flash_attention import _flash_attention_bwd_dq, _flash_attention_bwd_dkv
 
-    q, k, v, o, l, m, q_segment_ids, kv_segment_ids, ab = ctx.saved_tensors
+    q, k, v, o, l, m, q_segment_ids_fa, kv_segment_ids_fa, ab = ctx.saved_tensors
     causal = ctx.causal
     sm_scale = ctx.sm_scale
     partition_spec = ctx.partition_spec
     mesh = ctx.mesh
     full_shape = ctx.full_shape
+    # this segment_ids only reflects the local shape of segment_ids
     segment_ids = ctx.segment_ids
     grad_q = grad_k = grad_v = grad_ab = None
 
@@ -398,7 +409,7 @@ def backward(ctx, grad_output):
       if ab is not None:
         args += [ab]
       if segment_ids is not None:
-        args += [q_segment_ids, kv_segment_ids]
+        args += [q_segment_ids_fa, kv_segment_ids_fa]
       args += [expanded_l, expanded_m, grad_output, expanded_grad_i]
 
       outputs = [q]