Expanded vmap support for flash_mha. Vmapping q but not k,v reduces to a grouped-query attention, which we now support.

Author: nshepperd

Diff for: src/flash_attn_jax/flash.py

@@ -19,6 +19,7 @@
 from jax.sharding import PositionalSharding
 
 from einops import rearrange
+import einops
 import math
 
 from .flash_sharding import _flash_mha_fwd_hlo_sharded, _flash_mha_bwd_hlo_sharded
@@ -104,28 +105,91 @@ def _flash_mha_bwd_abstract(dout, q, k, v, out, lse, softmax_scale=None, is_caus
 # ==== VMap rules ====
 
 def mha_fwd_batch(vector_arg_values, batch_axes, **kwargs):
-  assert tuple(batch_axes) == (0,0,0), "Only support vmapping mha over axis 0 for now,"
-  [q, k, v] = vector_arg_values
-  [b, n, l, h, d] = q.shape
-  [b, n, lk, hk, d] = k.shape
-  assert [b, n, lk, hk, d] == list(v.shape)
-  out, lse = _flash_mha_fwd_p.bind(q.reshape([b*n,l,h,d]),
-                                   k.reshape([b*n,lk,hk,d]),
-                                   v.reshape([b*n,lk,hk,d]),
-                                   **kwargs)
-  return (out.reshape([b,n,*out.shape[1:]]), lse.reshape([b,n,*lse.shape[1:]])), (0,0)
+  assert all(isinstance(b, int) or b is None for b in batch_axes)
+  mapped = tuple(isinstance(b, int) for b in batch_axes)
+  if mapped == (True, True, True):
+    x = vector_arg_values[0].shape[batch_axes[0]]
+    def squish(val, axis):
+        dims = ['n', 'l', 'h', 'd']
+        dims.insert(axis, 'x')
+        dims = ' '.join(dims)
+        return einops.rearrange(val, f'{dims} -> (x n) l h d')
+    def unsquish(val):
+        return einops.rearrange(val, f'(x n) ... -> x n ...', x=x)
+    [q, k, v] = [squish(x, axis) for x, axis in zip(vector_arg_values, batch_axes)]
+    out, lse = _flash_mha_fwd_p.bind(q, k, v, **kwargs)
+    return (unsquish(out), unsquish(lse)), (0,0)
+  elif mapped == (True, False, False):
+    # This is just a GQA!
+    x = vector_arg_values[0].shape[batch_axes[0]]
+    def squish(val, axis):
+        if axis is None:
+            return val
+        dims = ['n', 'l', 'h', 'd']
+        dims.insert(axis, 'x')
+        dims = ' '.join(dims)
+        return einops.rearrange(val, f'{dims} -> n l (h x) d')
+    def unsquish(val):
+        return einops.rearrange(val, 'n l (h x) d -> x n l h d', x=x)
+    [q, k, v] = [squish(x, axis) for x, axis in zip(vector_arg_values, batch_axes)]
+    out, lse = _flash_mha_fwd_p.bind(q, k, v, **kwargs)
+    out = einops.rearrange(out, 'n l (h x) d -> x n l h d', x=x)
+    lse = einops.rearrange(lse, 'n (h x) l -> x n h l', x=x)
+    return (out, lse), (0,0)
+  else:
+    raise NotImplementedError("MHA fwd only support vmapping over q or (q,k,v) for now, got batch axes " + str(batch_axes))
 
 def mha_bwd_batch(vector_arg_values, batch_axes, **kwargs):
-  assert tuple(batch_axes) == (0,0,0,0,0,0), "Only support vmapping mha over axis 0 for now,"
-  dout, q, k, v, out, lse = vector_arg_values
-  b = dout.shape[batch_axes[0]]
-  def join(*args):
-      return [rearrange(a, 'b n ... -> (b n) ...') for a in args]
-  def unjoin(*args):
-      return [rearrange(a, '(b n) ... -> b n ...', b=b) for a in args]
-  dq, dk, dv = _flash_mha_bwd_p.bind(*join(dout,q,k,v,out,lse),
-                                     **kwargs)
-  return tuple(unjoin(dq,dk,dv)), (0,0,0)
+    assert all(isinstance(b, int) or b is None for b in batch_axes)
+    mapped = tuple(isinstance(b, int) for b in batch_axes)
+    if mapped == (True, True, True, True, True, True):
+        x = vector_arg_values[0].shape[batch_axes[0]]
+        def squish(val, axis):
+            if len(val.shape) == 5:
+                # q/k/v/o
+                dims = ['n', 'l', 'h', 'd']
+                dims.insert(axis, 'x')
+                dims = ' '.join(dims)
+                return einops.rearrange(val, f'{dims} -> (x n) l h d')
+            elif len(val.shape) == 4:
+                # lse
+                dims = ['n', 'h', 'l']
+                dims.insert(axis, 'x')
+                dims = ' '.join(dims)
+                return einops.rearrange(val, f'{dims} -> (x n) h l')
+        do, q, k, v, o, lse = [squish(x, axis) for x, axis in zip(vector_arg_values, batch_axes)]
+        dq, dk, dv = _flash_mha_bwd_p.bind(do, q, k, v, o, lse, **kwargs)
+        dq = einops.rearrange(dq, '(n x) l h d -> x n l h d', x=x)
+        dk = einops.rearrange(dk, '(n x) l h d -> x n l h d', x=x)
+        dv = einops.rearrange(dv, '(n x) l h d -> x n l h d', x=x)
+        return (dq,dk,dv), (0,0,0)
+    elif mapped == (True, True, False, False, True, True):
+        # Everything is mapped except k and v, which is a GQA backward
+        x = vector_arg_values[0].shape[batch_axes[0]]
+        def squish(val, axis):
+            if len(val.shape) == 5:
+                # q/k/v/o
+                dims = ['n', 'l', 'h', 'd']
+                dims.insert(axis, 'x')
+                dims = ' '.join(dims)
+                return einops.rearrange(val, f'{dims} -> n l (h x) d')
+            elif len(val.shape) == 4:
+                # lse
+                dims = ['n', 'h', 'l']
+                dims.insert(axis, 'x')
+                dims = ' '.join(dims)
+                return einops.rearrange(val, f'{dims} -> n (h x) l')
+        do = squish(vector_arg_values[0], batch_axes[0])
+        q = squish(vector_arg_values[1], batch_axes[1])
+        k = vector_arg_values[2]
+        v = vector_arg_values[3]
+        o = squish(vector_arg_values[4], batch_axes[4])
+        lse = squish(vector_arg_values[5], batch_axes[5])
+        dq, dk, dv = _flash_mha_bwd_p.bind(do, q, k, v, o, lse, **kwargs)
+        dq = einops.rearrange(dq, 'n l (h x) d -> x n l h d', x=x)
+        return (dq,dk,dv), (0,None,None)
+    else:
+        raise NotImplementedError("MHA bwd only support vmapping over q or (q,k,v) for now, got batch axes " + str(batch_axes))
 
 batching.primitive_batchers[_flash_mha_fwd_p] = mha_fwd_batch
 batching.primitive_batchers[_flash_mha_bwd_p] = mha_bwd_batch

Diff for: tests/test_flash.py

@@ -4,6 +4,7 @@
     sys.path.insert(0, glob.glob('build/lib.linux-*')[0])
 sys.path.insert(0,'./src')
 
+from functools import partial
 import pytest
 import jax
 import jax.numpy as jnp
@@ -100,22 +101,114 @@ def test_flash_fwd_vmap(n, seqlen, h, d, causal, local, dtype):
     k = jax.random.normal(jax.random.PRNGKey(1), [x, n, seqlen, h, d], dtype=jnp.float32)
     v = jax.random.normal(jax.random.PRNGKey(2), [x, n, seqlen, h, d], dtype=jnp.float32)
 
-    @jax.jit
     def ref(q,k,v):
         return ref_mha(q,k,v, is_causal=bool(causal), window_size=window_size)
-    @jax.jit
     def flash(q,k,v):
         return flash_mha(q,k,v, is_causal=bool(causal), window_size=window_size)
 
-    ref_out = jnp.stack([ref(q[i],k[i],v[i]) for i in range(x)])
+    ref_out = jax.vmap(ref)(q,k,v)
     q = q.astype(dtype)
     k = k.astype(dtype)
     v = v.astype(dtype)
-    f16_out = jnp.stack([ref(q[i],k[i],v[i]) for i in range(x)])
-
+    f16_out = jax.vmap(ref)(q,k,v)
 
     out = jax.vmap(flash)(q,k,v)
     check(ref_out, f16_out, out)
 
+@pytest.mark.parametrize("dtype", [jnp.float16, jnp.bfloat16])
+@pytest.mark.parametrize("local", ['local',''])
+@pytest.mark.parametrize("causal", ['causal',''])
+@pytest.mark.parametrize("d", [59, 32])
+@pytest.mark.parametrize("h", [1, 4])
+@pytest.mark.parametrize("seqlen", [97, 128])
+@pytest.mark.parametrize("n", [1])
+def test_flash_fwd_vmapq(n, seqlen, h, d, causal, local, dtype):
+    window_size = (3,3) if local else (-1,-1)
+
+    x = 4
+    q = jax.random.normal(jax.random.PRNGKey(0), [x, n, seqlen, h, d], dtype=jnp.float32)
+    k = jax.random.normal(jax.random.PRNGKey(1), [n, seqlen, h, d], dtype=jnp.float32)
+    v = jax.random.normal(jax.random.PRNGKey(2), [n, seqlen, h, d], dtype=jnp.float32)
+
+    def ref(q,k,v):
+        return ref_mha(q,k,v, is_causal=bool(causal), window_size=window_size)
+    def flash(q,k,v):
+        return flash_mha(q,k,v, is_causal=bool(causal), window_size=window_size)
+
+    ref_out = jax.vmap(ref, in_axes=(0,None,None))(q,k,v)
+    q = q.astype(dtype)
+    k = k.astype(dtype)
+    v = v.astype(dtype)
+    f16_out = jax.vmap(ref, in_axes=(0,None,None))(q,k,v)
+
+    out = jax.vmap(flash, in_axes=(0,None,None))(q,k,v)
+    check(ref_out, f16_out, out)
+
+@pytest.mark.parametrize("dtype", [jnp.float16, jnp.bfloat16])
+@pytest.mark.parametrize("local", ['local',''])
+@pytest.mark.parametrize("causal", ['causal',''])
+@pytest.mark.parametrize("d", [59, 32])
+@pytest.mark.parametrize("h", [1, 4])
+@pytest.mark.parametrize("seqlen", [97, 128])
+@pytest.mark.parametrize("n", [1])
+def test_flash_bwd_vmap(n, seqlen, h, d, causal, local, dtype):
+    window_size = (3,3) if local else (-1,-1)
+
+    x = 4
+    q = jax.random.normal(jax.random.PRNGKey(0), [x, n, seqlen, h, d], dtype=jnp.float32)
+    k = jax.random.normal(jax.random.PRNGKey(1), [x, n, seqlen, h, d], dtype=jnp.float32)
+    v = jax.random.normal(jax.random.PRNGKey(2), [x, n, seqlen, h, d], dtype=jnp.float32)
+    do = jax.random.normal(jax.random.PRNGKey(3), [x, n, seqlen, h, d], dtype=jnp.float32)
+
+    def func(mha, q,k,v):
+        @partial(jax.vmap, in_axes=(0,0,0))
+        def fwd(q,k,v):
+            return mha(q,k,v, is_causal=bool(causal), window_size=window_size)
+        o, bwd = jax.vjp(fwd,q,k,v)
+        return bwd(do)
+
+    ref_out = func(ref_mha, q,k,v)
+    q = q.astype(dtype)
+    k = k.astype(dtype)
+    v = v.astype(dtype)
+    do = do.astype(dtype)
+    f16_out = func(ref_mha, q,k,v)
+
+    out = func(flash_mha, q,k,v)
+    check(ref_out, f16_out, out)
+
+@pytest.mark.parametrize("dtype", [jnp.float16, jnp.bfloat16])
+@pytest.mark.parametrize("local", ['local',''])
+@pytest.mark.parametrize("causal", ['causal',''])
+@pytest.mark.parametrize("d", [59, 32])
+@pytest.mark.parametrize("h", [1, 4])
+@pytest.mark.parametrize("seqlen", [97, 128])
+@pytest.mark.parametrize("n", [1])
+def test_flash_bwd_vmapq(n, seqlen, h, d, causal, local, dtype):
+    window_size = (3,3) if local else (-1,-1)
+
+    x = 4
+    q = jax.random.normal(jax.random.PRNGKey(0), [x, n, seqlen, h, d], dtype=jnp.float32)
+    k = jax.random.normal(jax.random.PRNGKey(1), [n, seqlen, h, d], dtype=jnp.float32)
+    v = jax.random.normal(jax.random.PRNGKey(2), [n, seqlen, h, d], dtype=jnp.float32)
+    do = jax.random.normal(jax.random.PRNGKey(3), [x, n, seqlen, h, d], dtype=jnp.float32)
+
+    def func(mha, q,k,v):
+        @partial(jax.vmap, in_axes=(0,None,None))
+        def fwd(q,k,v):
+            return mha(q,k,v, is_causal=bool(causal), window_size=window_size)
+        o, bwd = jax.vjp(fwd,q,k,v)
+        return bwd(do)
+
+    ref_out = func(ref_mha, q,k,v)
+    q = q.astype(dtype)
+    k = k.astype(dtype)
+    v = v.astype(dtype)
+    do = do.astype(dtype)
+    f16_out = func(ref_mha, q,k,v)
+
+    out = func(flash_mha, q,k,v)
+    check(ref_out, f16_out, out)
+
 if __name__ == '__main__':
     test_flash_bwd(1,4,1,32,4,False,False,jnp.float16)