Skip to content
The Chat Archive
Chatbot Arena (model unspecified) ShareLM (browser-extension contributions)

You will recieve two files, one of them is a modification of the other, please note all…

  1. User

    You will recieve two files, one of them is a modification of the other, please note all the changes.File 1:

    """Flash Attention TPU kernel.""" from future import annotations

    import dataclasses import functools from typing import Any, NamedTuple

    import jax from jax import lax from jax.experimental import pallas as pl from jax.experimental.pallas import tpu as pltpu import jax.numpy as jnp

    DEFAULT_MASK_VALUE = -0.7 * float(jnp.finfo(jnp.dtype("float32")).max) NUM_LANES = 128 NUM_SUBLANES = 8

    class SegmentIds(NamedTuple): """SegmentIds for Q and KV sequences.

    SegmentIds are used to generate segment mask, which prevents attention between different segments in the input sequence. Each array is a list of ids (integers). Only the token with the same id can attend to each other.

    Attributes: q: segment ids along the Q sequence. kv: segment ids along the KV sequence. """

    q: jax.Array # [batch_size, q_seq_len] kv: jax.Array # [batch_size, kv_seq_len]

    @dataclasses.dataclass(frozen=True) class BlockSizes: """Tile sizes parameterizing FlashAttention kernels.

    Those parameters have negligible effect on numerics, but affect performance greatly. """ block_q: int block_k_major: int block_k: int block_b: int

    block_q_major_dkv: int | None = None block_k_major_dkv: int | None = None block_k_dkv: int | None = None block_q_dkv: int | None = None

    block_k_major_dq: int | None = None block_k_dq: int | None = None block_q_dq: int | None = None

    def post_init(self): def verify_major_minor(prefix, suffix, major, minor): if minor > major: raise ValueError( f"{prefix}{suffix}={minor} should be smaller than" f" {prefix}_major{suffix}={major}" ) if major % minor != 0: raise ValueError( f"{prefix}{suffix}={minor} should divide" f" {prefix}_major{suffix}={major}" )

    verify_major_minor("block_k", "", self.block_k_major, self.block_k)
    if self.block_q_major_dkv is not None and self.block_q_dkv is not None:
      verify_major_minor(
          "block_q", "_dkv", self.block_q_major_dkv, self.block_q_dkv
      )
    if self.block_k_major_dkv is not None and self.block_k_dkv is not None:
      verify_major_minor(
          "block_k", "_dkv", self.block_k_major_dkv, self.block_k_dkv
      )
    if self.block_k_major_dq is not None and self.block_k_dq is not None:
      verify_major_minor(
          "block_k", "_dq", self.block_k_major_dq, self.block_k_dq
      )
    

    @property def has_backward_blocks(self) -> bool: backward_blocks = ( self.block_q_major_dkv, self.block_k_major_dkv, self.block_q_dkv, self.block_k_dkv, self.block_k_major_dq, self.block_k_dq, self.block_q_dq, ) return all(b is not None for b in backward_blocks)

    @classmethod def get_default(cls, batch_size, num_heads, q_seq_len, kv_len, d_model): # TODO(apaszke,sharadmv): Select better parameters based on a heuristic. del batch_size, num_heads, q_seq_len, kv_len, d_model # Unused. return BlockSizes( block_q=128, block_k_major=128, block_k=128, block_b=1, block_q_major_dkv=128, block_k_major_dkv=128, block_k_dkv=128, block_q_dkv=128, block_k_major_dq=128, block_k_dq=128, block_q_dq=128, )

    @functools.partial( jax.jit, static_argnames=[ "causal", "sm_scale", "block_sizes", "debug", ], ) def flash_attention( q, # [batch_size, num_heads, q_seq_len, d_model] k, # [batch_size, num_heads, kv_seq_len, d_model] v, # [batch_size, num_heads, kv_seq_len, d_model] ab=None, # [batch_size, num_heads, q_seq_len, kv_seq_len] segment_ids=None, # q of [batch_size, q_seq_len] and kv of [batch_size, kv_seq_len] *, causal: bool = False, sm_scale: float = 1.0, block_sizes: BlockSizes | None = None, debug: bool = False, ): batch_size, num_heads, q_seq_len, d_model = q.shape batch_size_k, num_heads_k, kv_seq_len, d_model_k = k.shape batch_size_v, num_heads_v, kv_seq_len_v, d_model_v = v.shape if batch_size != batch_size_k or batch_size != batch_size_v: raise ValueError( f"Batch size mismatch: got {batch_size}, {batch_size_k} and" f" {batch_size_v} (for q, k, v respectively)" ) if num_heads != num_heads_k or num_heads != num_heads_v: raise ValueError( f"Head count mismatch: got {num_heads}, {num_heads_k}," f" {num_heads_v} (for q, k, v respectively)" ) if d_model != d_model_k: raise ValueError( f"Model dimension mismatch: got {d_model} and {d_model_k} (for q and k" " respectively)" ) if d_model != d_model_v: raise NotImplementedError( "V model dimension unequal to KV model dimension unsupported" ) if kv_seq_len != kv_seq_len_v: raise ValueError( f"KV sequence length mismatch: got {kv_seq_len} and {kv_seq_len_v}" ) if ab is not None: if ab.shape != (batch_size, num_heads, q_seq_len, kv_seq_len): raise ValueError( f"Attention bias shape mismatch: expected ({batch_size=}," f" {num_heads=}, {q_seq_len=}, {kv_seq_len=}), got {ab.shape}" ) if segment_ids is not None: if segment_ids.q.shape != (batch_size, q_seq_len): raise ValueError( f"Q segment ids shape mismatch: expected ({batch_size=}," f" {q_seq_len=},), got {segment_ids.q.shape}" ) if segment_ids.kv.shape != (batch_size, kv_seq_len): raise ValueError( f"KV segment ids shape mismatch: expected ({batch_size=}," f" {kv_seq_len=},), got {segment_ids.kv.shape}" ) if block_sizes is None: block_sizes = BlockSizes.get_default( batch_size, num_heads, q_seq_len, kv_seq_len, d_model ) return _flash_attention( q, k, v, ab, segment_ids, False, causal, sm_scale, block_sizes, debug )

    @functools.partial(jax.custom_vjp, nondiff_argnums=range(5, 10)) def _flash_attention( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_sizes, debug, ): return _flash_attention_impl( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_sizes.block_b, block_sizes.block_q, block_sizes.block_k_major, block_sizes.block_k, debug, )

    def _flash_attention_fwd( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_sizes, debug, ): if save_residuals: raise NotImplementedError("Higher-order AD not supported") o, l, m = _flash_attention( q, k, v, ab, segment_ids, True, causal, sm_scale, block_sizes, debug ) return o, (q, k, v, ab, segment_ids, o, l, m)

    def _flash_attention_bwd( save_residuals: bool, causal: bool, sm_scale: float, block_sizes: BlockSizes, debug: bool, residuals, do, ): """VJP rule for FlashAttention.""" if save_residuals: raise NotImplementedError("Higher-order AD not supported") (q, k, v, ab, segment_ids, o, l, m) = residuals if not block_sizes.has_backward_blocks: raise ValueError( "Program is being differentiated, but not all backward blocks are" " specified" )

    di = jnp.sum( o.astype(jnp.float32) * do.astype(jnp.float32), axis=-1 ) # [batch_size, num_heads, q_seq_len]

    dk, dv = _flash_attention_bwd_dkv( q, k, v, ab, segment_ids, l, m, do, di, block_q_major=block_sizes.block_q_major_dkv, block_k_major=block_sizes.block_k_major_dkv, block_k=block_sizes.block_k_dkv, block_q=block_sizes.block_q_dkv, sm_scale=sm_scale, causal=causal, mask_value=DEFAULT_MASK_VALUE, debug=debug, )

    dq, ds = _flash_attention_bwd_dq( q, k, v, ab, segment_ids, l, m, do, di, block_q_major=block_sizes.block_q_dq, block_k_major=block_sizes.block_k_major_dq, block_k=block_sizes.block_k_dq, sm_scale=sm_scale, causal=causal, mask_value=DEFAULT_MASK_VALUE, debug=debug, ) return dq, dk, dv, ds, None

    _flash_attention.defvjp(fwd=_flash_attention_fwd, bwd=_flash_attention_bwd)

    MIN_BLOCK_SIZE = 128 TRANS_B_DIM_NUMBERS = (((1,), (1,)), ((), ()))

    def below_or_on_diag(r, r_blk_size, c, c_blk_size):

    A block is considered below or on diagonal as long as the bottom left

    corner of the block is below or on diagonal.

    return ((r + 1) * r_blk_size - 1) > (c * c_blk_size)

    def _flash_attention_kernel(q_tile_ref, *args, **kwargs): block_b = q_tile_ref.shape[0]

    If we're not going to tile the softmax, then we can avoid a bunch of VPU ops.

    if kwargs["block_k"] == kwargs["kv_seq_len"]: kernel = _flash_attention_kernel_single_batch_single_step else: kernel = _flash_attention_kernel_single_batch for batch_idx in range(block_b): kernel((batch_idx, 0), q_tile_ref, *args, **kwargs)

    def _flash_attention_kernel_single_batch( batch_idx: tuple[int, ...], q_tile_ref, k_tile_ref, v_tile_ref, ab_tile_ref, q_segment_ids_tile_ref, kv_segment_ids_tile_ref, # Input arrays o_tile_ref, # Output arrays l_ref, m_ref, m_scratch_ref, l_scratch_ref, acc_scratch_ref, *, causal, sm_scale, block_k, kv_seq_len, mask_value, ): block_k_major = k_tile_ref.shape[2] block_q = q_tile_ref.shape[2] head_dim = q_tile_ref.shape[-1]

    kv_seq_idx = pl.program_id(3) @pl.when(kv_seq_idx == 0) def start_new_sequence(): m_scratch_ref[batch_idx] = jnp.full( m_scratch_ref.shape[2:], -jnp.inf, jnp.float32 ) l_scratch_ref[batch_idx] = jnp.zeros(l_scratch_ref.shape[2:], jnp.float32) acc_scratch_ref[batch_idx] = jnp.zeros( acc_scratch_ref.shape[2:], jnp.float32 )

    q_seq_idx = pl.program_id(2) if causal: should_run = below_or_on_diag(q_seq_idx, block_q, kv_seq_idx, block_k_major) else: should_run = True

    @pl.when(should_run) def run(): @functools.partial( lax.fori_loop, 0, block_k_major // block_k, init_val=None, unroll=True ) def body(i, _): m_prev = m_scratch_ref[batch_idx] l_prev = l_scratch_ref[batch_idx] q = q_tile_ref[batch_idx] # [block_q, head_dim] start_k = i * block_k k = pl.load( k_tile_ref, (*batch_idx, pl.dslice(start_k, block_k), slice(None)) ) # [block_k, head_dim]

      s = jax.lax.dot_general(
          q, k, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32
      )  # [block_q, block_k]
    
      # Add attention bias if needed.
      # TODO(tanburn) Should the attention bias be added before or after
      # multiplication by sm_scale?
      if ab_tile_ref is not None:
        ab = pl.load(
            ab_tile_ref,
            (*batch_idx, pl.dslice(None), pl.dslice(start_k, block_k))
        ).astype(jnp.float32)
        s += ab
    
      if sm_scale != 1.0:
        s *= sm_scale
    
      mask = None
      if q_segment_ids_tile_ref is not None:
        repeats, rem = divmod(block_k, NUM_LANES)
        if rem:
          raise NotImplementedError(
              f"kv block size must be a multiple of {NUM_LANES}"
          )
        q_segment_ids = pltpu.repeat(
            q_segment_ids_tile_ref[batch_idx[0]], repeats, axis=1
        )  # [block_q, block_k].
        kv_segment_ids = pl.load(
            kv_segment_ids_tile_ref,
            (batch_idx[0], pl.dslice(1), pl.dslice(start_k, block_k)),
        )  # [1, block_k].
        mask = jnp.equal(q_segment_ids, kv_segment_ids).astype(jnp.bool_)
    
      if causal:
        mask_shape = (block_q, block_k)
        row_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 0)
        row_ids += q_seq_idx * block_q
        col_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 1)
        col_ids += kv_seq_idx * block_k_major + start_k
        causal_mask = col_ids <= row_ids
        mask = (
            causal_mask if mask is None else jnp.logical_and(mask, causal_mask)
        )
    
      s = s if mask is None else s + jnp.where(mask, 0.0, mask_value)
    
      m_curr = jnp.max(s, axis=1)[:, None]  # Row max, shape [block_q, 1].
      m_next = jnp.maximum(m_prev, m_curr)  # Shape [block_q, 128].
    
      block_k_repeats, rem = divmod(block_k, MIN_BLOCK_SIZE)
      if rem:
        raise NotImplementedError(
            f"{block_k=} should be a multiple of {MIN_BLOCK_SIZE}"
        )
      p = jnp.exp(s - pltpu.repeat(m_next, block_k_repeats, 1))
    
      alpha = jnp.exp(m_prev - m_next)  # Shape [block_q, 128].
    
      l_corr = alpha * l_prev
    
      l_next = jnp.sum(p, axis=1)[:, None] + l_corr  # Shape [block_q, 128]
    
      head_dim_repeats, rem = divmod(head_dim, MIN_BLOCK_SIZE)
      l_broadcast = lambda l: pltpu.repeat(l, head_dim_repeats, 1)
      if rem:
        if head_dim_repeats == 0:
          l_broadcast = lambda l: l[:, :head_dim]
        else:
          raise NotImplementedError(
              f"{head_dim=} should be a multiple of {MIN_BLOCK_SIZE} if larger"
          )
      l_scratch_ref[batch_idx] = l_next
      m_scratch_ref[batch_idx] = m_next
    
      l_next_inv_safe = jnp.where(l_next == 0.0, 1.0, 1.0 / l_next)
      acc_scratch_ref[batch_idx] *= l_broadcast(l_corr * l_next_inv_safe)
      v = pl.load(
          v_tile_ref, (*batch_idx, pl.dslice(start_k, block_k), slice(None))
      )
      o_curr = jax.lax.dot(
          p.astype(v.dtype), v, preferred_element_type=jnp.float32
      )
      acc_scratch_ref[batch_idx] += o_curr * l_broadcast(l_next_inv_safe)
    

    @pl.when(kv_seq_idx == (kv_seq_len // block_k_major) - 1) def store_output(): o_tile_ref[batch_idx] = acc_scratch_ref[batch_idx].astype(o_tile_ref.dtype) if l_ref is not None: l_ref[batch_idx] = l_scratch_ref[batch_idx].astype(l_ref.dtype) if m_ref is not None: m_ref[batch_idx] = m_scratch_ref[batch_idx].astype(m_ref.dtype)

    def _flash_attention_kernel_single_batch_single_step( batch_idx: tuple[int, ...], q_tile_ref, k_tile_ref, v_tile_ref, ab_tile_ref, q_segment_ids_tile_ref, kv_segment_ids_tile_ref, # Input arrays o_tile_ref, # Output arrays l_ref: Any | None = None, m_ref: Any | None = None, *, causal, sm_scale, block_k, kv_seq_len, mask_value, ): block_k_major = k_tile_ref.shape[2] block_q = q_tile_ref.shape[2]

    assert kv_seq_len == block_k_major == block_k

    q = q_tile_ref[batch_idx] # [block_q, head_dim] k = k_tile_ref[batch_idx] # [block_k, head_dim] s = jax.lax.dot_general( q, k, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32 ) # [block_q, block_k]

    if ab_tile_ref is not None: s += ab_tile_ref[batch_idx].astype(jnp.float32) if sm_scale != 1.0: s *= sm_scale

    mask = None if q_segment_ids_tile_ref is not None: repeats, rem = divmod(block_k, NUM_LANES) if rem: raise NotImplementedError( f"kv block size must be a multiple of {NUM_LANES}" ) q_segment_ids = pl.load( q_segment_ids_tile_ref, (batch_idx[0],) ) # [block_q, NUM_LANES]. q_segment_ids = pltpu.repeat( q_segment_ids, repeats, axis=1 ) # [block_q, block_k]. kv_segment_ids = pl.load( kv_segment_ids_tile_ref, (batch_idx[0], pl.dslice(1)) ) # [1, block_k]. mask = jnp.equal(q_segment_ids, kv_segment_ids).astype(jnp.bool_)

    if causal: q_seq_idx = pl.program_id(2) mask_shape = (block_q, block_k) row_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 0) row_ids += q_seq_idx * block_q col_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 1) causal_mask = col_ids <= row_ids mask = causal_mask if mask is None else jnp.logical_and(mask, causal_mask) s = s if mask is None else s + jnp.where(mask, 0.0, mask_value)

    m = jnp.max(s, axis=1)[:, None] p = jnp.exp(s - m) l = jnp.sum(p, axis=1)[:, None] p /= l

    if m_ref is not None: m_ref[batch_idx] = lax.broadcast_in_dim(m, m_ref.shape[2:], range(2)) if l_ref is not None: l_ref[batch_idx] = lax.broadcast_in_dim(l, l_ref.shape[2:], range(2))

    v = v_tile_ref[batch_idx] o_tile_ref[batch_idx] = jax.lax.dot( p.astype(v.dtype), v, preferred_element_type=jnp.float32 ).astype(o_tile_ref.dtype)

    def _flash_attention_impl( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_b, block_q, block_k_major, block_k, debug, ): batch_size, num_heads, q_seq_len, head_dim = q.shape _, _, kv_seq_len, _ = k.shape _verify_block("block_q", "q_seq_len", block_q, q_seq_len, should_divide=False) _verify_block("block_k_major", "kv_seq_len", block_k_major, kv_seq_len) _verify_block("block_k", "kv_seq_len", block_k, kv_seq_len) _verify_block("block_b", "batch", block_b, batch_size, should_divide=False)

    TODO(apaszke): Tile over heads as well.

    grid = ( pl.cdiv(batch_size, block_b), num_heads, pl.cdiv(q_seq_len, block_q), kv_seq_len // block_k_major, )

    def q_index_map(batch_index, head_index, q_seq_index, _): return (batch_index, head_index, q_seq_index, 0)

    def kv_index_map(batch_index, head_index, q_seq_index, kv_seq_index): if causal: # If the kv block is skipped, prefetch the next valid kv block, i.e. the # 0th one to be used for the next block_q rows. next_kv_index = lax.select( below_or_on_diag(q_seq_index, block_q, kv_seq_index, block_k_major), kv_seq_index, 0, ) else: next_kv_index = kv_seq_index return (batch_index, head_index, next_kv_index, 0)

    def ab_index_map(batch_index, head_index, q_seq_index, kv_seq_index): if causal: should_run = below_or_on_diag( q_seq_index, block_q, kv_seq_index, block_k_major ) # If the ab block is skipped, prefetch the next valid ab block, i.e. the # 0th kv to be used for the next block_q rows. next_q_index = lax.select( should_run, q_seq_index, lax.select( q_seq_index == (q_seq_len // block_q) - 1, 0, q_seq_index + 1 ), ) next_kv_index = lax.select(should_run, kv_seq_index, 0) else: next_q_index = q_seq_index next_kv_index = kv_seq_index

    return (batch_index, head_index, next_q_index, next_kv_index)
    

    def o_index_map(batch_index, head_index, q_seq_index, _): return (batch_index, head_index, q_seq_index, 0)

    def lm_index_map(batch_index, head_index, q_seq_index, _): return (batch_index, head_index, q_seq_index, 0)

    kernel = functools.partial( _flash_attention_kernel, causal=causal, mask_value=DEFAULT_MASK_VALUE, sm_scale=sm_scale, block_k=block_k, kv_seq_len=kv_seq_len, ) out_shape = jax.ShapeDtypeStruct(shape=q.shape, dtype=q.dtype) out_shape = [out_shape] out_specs = [pl.BlockSpec((block_b, 1, block_q, head_dim), o_index_map)]

    if block_k != kv_seq_len: m_scratch = pltpu.VMEM((block_b, 1, block_q, MIN_BLOCK_SIZE), jnp.float32) l_scratch = pltpu.VMEM((block_b, 1, block_q, MIN_BLOCK_SIZE), jnp.float32) acc_scratch = pltpu.VMEM((block_b, 1, block_q, head_dim), jnp.float32) scratch_shapes = [m_scratch, l_scratch, acc_scratch] else: scratch_shapes = []

    if save_residuals: out_specs = [ *out_specs, pl.BlockSpec((block_b, 1, block_q, MIN_BLOCK_SIZE), lm_index_map), pl.BlockSpec((block_b, 1, block_q, MIN_BLOCK_SIZE), lm_index_map), ] l = jax.ShapeDtypeStruct( (batch_size, num_heads, q_seq_len, MIN_BLOCK_SIZE), dtype=jnp.float32 ) m = jax.ShapeDtypeStruct( (batch_size, num_heads, q_seq_len, MIN_BLOCK_SIZE), dtype=jnp.float32 ) out_shape = (*out_shape, l, m) else: out_specs = [*out_specs, None, None] out_shape = (*out_shape, None, None)

    ab_block_spec = ( pl.BlockSpec((block_b, 1, block_q, block_k_major), ab_index_map) if ab is not None else None)

    q_segment_ids_spec = kv_segment_ids_spec = None q_segment_ids = kv_segment_ids = None if segment_ids is not None:

    def q_segment_ids_index_map(batch_index, head_index, q_seq_index, _):
      del head_index
      return (batch_index, q_seq_index, 0)
    
    def kv_segment_ids_index_map(
        batch_index, head_index, q_seq_index, kv_seq_index
    ):
      del head_index
      if causal:
        next_kv_index = lax.select(
            below_or_on_diag(q_seq_index, block_q, kv_seq_index, block_k_major),
            kv_seq_index,
            0,
        )
      else:
        next_kv_index = kv_seq_index
      return (batch_index, 0, next_kv_index)
    
    q_segment_ids_spec = pl.BlockSpec(
        (block_b, block_q, NUM_LANES), q_segment_ids_index_map
    )
    kv_segment_ids_spec = pl.BlockSpec(
        (block_b, NUM_SUBLANES, block_k_major), kv_segment_ids_index_map
    )
    
    q_segment_ids = jax.lax.broadcast_in_dim(
        segment_ids.q,
        (batch_size, q_seq_len, NUM_LANES),
        (
            0,
            1,
        ),
    )
    kv_segment_ids = jax.lax.broadcast_in_dim(
        segment_ids.kv,
        (batch_size, NUM_SUBLANES, kv_seq_len),
        (
            0,
            2,
        ),
    )
    

    in_specs = [ pl.BlockSpec((block_b, 1, block_q, head_dim), q_index_map), pl.BlockSpec((block_b, 1, block_k_major, head_dim), kv_index_map), pl.BlockSpec((block_b, 1, block_k_major, head_dim), kv_index_map), ab_block_spec, q_segment_ids_spec, kv_segment_ids_spec, ]

    o, *aux = pl.pallas_call( kernel, grid_spec=pltpu.PrefetchScalarGridSpec( num_scalar_prefetch=0, grid=grid, in_specs=in_specs, out_specs=out_specs, scratch_shapes=scratch_shapes, ), out_shape=out_shape, debug=debug, compiler_params=dict( mosaic=dict( dimension_semantics=( "parallel", "parallel", "parallel", "arbitrary", ) ) ), )(q, k, v, ab, q_segment_ids, kv_segment_ids) if save_residuals: l, m = (v[..., 0] for v in aux[-2:]) return (o, l, m) else: return o

    def _flash_attention_dkv_kernel( q_tile_ref, k_tile_ref, v_tile_ref, ab_tile_ref, q_segment_ids_tile_ref, kv_segment_ids_tile_ref, l_tile_ref, m_tile_ref, do_tile_ref, di_tile_ref, dk_tile_ref, dv_tile_ref, dk_scratch_ref, dv_scratch_ref, *, sm_scale: float, causal: bool, mask_value: float, q_seq_len: int, block_q: int, block_k: int, ): _, _, block_q_major, _ = q_tile_ref.shape _, _, block_k_major, _ = k_tile_ref.shape

    q_seq_index = pl.program_id(axis=3) kv_seq_index = pl.program_id(axis=2)

    @pl.when(q_seq_index == 0) def start_new_sequence(): dk_scratch_ref[:, :] = jnp.zeros(dk_scratch_ref.shape, dk_scratch_ref.dtype) dv_scratch_ref[:, :] = jnp.zeros(dv_scratch_ref.shape, dv_scratch_ref.dtype)

    def q_body(j, _): start_q = j * block_q def k_body(i, _): start_k = i * block_k k = pl.load(k_tile_ref, (0, 0, pl.ds(start_k, block_k), slice(None))) v = pl.load(v_tile_ref, (0, 0, pl.ds(start_k, block_k), slice(None))) q = pl.load(q_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, head_dim] l = pl.load(l_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, 128] m = pl.load(m_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, 128] do = pl.load(do_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, 128] di = pl.load(di_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ).astype(jnp.float32) # [block_q, 128]

      capped_logits = lax.dot_general(
          q, k, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32
      )  # [block_q_major, block_k]
    
      if ab_tile_ref is not None:
        ab = pl.load(
            ab_tile_ref,
            (
                0,
                0,
                pl.dslice(j * block_q, block_q),
                pl.dslice(i * block_k, block_k),
            ),
        ).astype(jnp.float32)
        capped_logits += ab
    
      if sm_scale != 1.0:
        capped_logits *= sm_scale
    
      mask = None
      if q_segment_ids_tile_ref is not None:
        repeats, rem = divmod(block_k, NUM_LANES)
        if rem:
          raise NotImplementedError(
          )
        q_segment_ids = pl.load(
            q_segment_ids_tile_ref, (0, pl.ds(start_q, block_q), slice(None))
        )  # [block_q, NUM_LANES].
        q_segment_ids = pltpu.repeat(
            q_segment_ids, repeats, axis=1
        )  # [block_q, block_k].
        kv_segment_ids = pl.load(
            kv_segment_ids_tile_ref, (slice(None), 0, pl.ds(start_k, block_k))
        )  # [1, block_k].
        mask = jnp.equal(q_segment_ids, kv_segment_ids).astype(jnp.bool_)
    
      if causal:
        mask_shape = (block_q, block_k)
        row_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 0)
        row_ids += q_seq_index * block_q_major + start_q
        col_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 1)
        col_ids += kv_seq_index * block_k_major + start_k
        causal_mask = col_ids <= row_ids
        mask = (
            causal_mask if mask is None else jnp.logical_and(mask, causal_mask)
        )
    
      capped_logits = (
          capped_logits
          if mask is None
          else capped_logits + jnp.where(mask, 0.0, mask_value)
      )
    
      p = jnp.exp(
          capped_logits - pltpu.repeat(m, block_k // MIN_BLOCK_SIZE, axis=1)
      )
      p = p * pltpu.repeat(
          1 / l, block_k // MIN_BLOCK_SIZE, axis=1
      )  # [block_q_major, block_k_major]
      dv = lax.dot(p.T.astype(do.dtype), do, preferred_element_type=jnp.float32)
      pl.store(dv_scratch_ref, (pl.ds(start_k, block_k), slice(None)),
               pl.load(dv_scratch_ref, (pl.ds(start_k, block_k), slice(None)))
               + dv.astype(dv_scratch_ref.dtype))
    
      # di: [block_q, 128]
      # do: [block_q, head_dim]
      # v: [block_k_major, head_dim]
      dp = lax.dot_general(
          do, v, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32
      )
      ds = (dp - pltpu.repeat(di, block_k // MIN_BLOCK_SIZE, axis=1)) * p
    
      if sm_scale != 1.0:
        ds = ds * sm_scale
    
      # ds: [block_q_major, block_k_major]
      # q: [block_q_major, head_dim]
      dk = lax.dot(ds.T.astype(do.dtype), q, preferred_element_type=jnp.float32)
      pl.store(dk_scratch_ref, (pl.ds(start_k, block_k), slice(None)),
               pl.load(dk_scratch_ref, (pl.ds(start_k, block_k), slice(None)))
               + dk.astype(dk_scratch_ref.dtype))
    lax.fori_loop(0, block_k_major // block_k, k_body, None, unroll=True)
    

    if causal: should_run = below_or_on_diag( q_seq_index, block_q_major, kv_seq_index, block_k_major ) else: should_run = True

    @pl.when(should_run) def run(): lax.fori_loop(0, block_q_major // block_q, q_body, None, unroll=True)

    @pl.when(q_seq_index == q_seq_len // block_q_major - 1) def end_of_q_sequence(): dv_tile_ref[0, 0, :, :] = dv_scratch_ref[...].astype(dv_tile_ref.dtype) dk_tile_ref[0, 0, :, :] = dk_scratch_ref[...].astype(dk_tile_ref.dtype)

    def _flash_attention_bwd_dkv( q, k, v, ab, segment_ids, l, m, do, di, *, block_q_major: int | None, block_q: int | None, block_k_major: int | None, block_k: int | None, sm_scale: float, causal: bool = False, mask_value: float = DEFAULT_MASK_VALUE, debug: bool = False, ): batch_size, num_heads, q_seq_len, head_dim = q.shape _, _, kv_seq_len, _ = k.shape _verify_block("block_q_major_dkv", "q_seq_len", block_q_major, q_seq_len) _verify_block("block_q_dkv", "q_seq_len", block_q, q_seq_len) _verify_block("block_k_major_dkv", "kv_seq_len", block_k_major, kv_seq_len) _verify_block("block_k_dkv", "kv_seq_len", block_k, kv_seq_len)

    Broadcast out scalar values

    m = jnp.broadcast_to(m[..., None], (*m.shape, MIN_BLOCK_SIZE)) l = jnp.broadcast_to(l[..., None], (*l.shape, MIN_BLOCK_SIZE))

    Preprocess contraction for bwd pass

    di = jnp.broadcast_to(di[..., None], (*di.shape, MIN_BLOCK_SIZE))

    kv index needs to be before q index since q index is the contractng

    dimension.

    grid = ( batch_size, num_heads, kv_seq_len // block_k_major, q_seq_len // block_q_major, )

    def qo_index_map(batch_index, head_index, kv_seq_index, q_seq_index): if causal: # If the q block is skipped, stay at the 0th q block. next_q_index = lax.select( below_or_on_diag( q_seq_index, block_q_major, kv_seq_index, block_k_major ), q_seq_index, 0, ) else: next_q_index = q_seq_index

    return (batch_index, head_index, next_q_index, 0)
    

    qo_spec = pl.BlockSpec((1, 1, block_q_major, head_dim), qo_index_map) assert qo_spec.block_shape is not None assert q.ndim == len(qo_spec.block_shape) do_spec = qo_spec assert do.ndim == len(qo_spec.block_shape)

    def kv_index_map(batch_index, head_index, kv_seq_index, _): return (batch_index, head_index, kv_seq_index, 0)

    kv_spec = pl.BlockSpec((1, 1, block_k_major, head_dim), kv_index_map) assert kv_spec.block_shape is not None assert k.ndim == len(kv_spec.block_shape) assert v.ndim == len(kv_spec.block_shape)

    def lm_index_map(batch_index, head_index, _, q_seq_index): return (batch_index, head_index, q_seq_index, 0)

    lm_spec = pl.BlockSpec((1, 1, block_q_major, MIN_BLOCK_SIZE), lm_index_

  2. Assistant

    File 2:

    """Flash Attention TPU kernel.""" from future import annotations

    import dataclasses import functools from typing import Any, NamedTuple

    import jax from jax import lax from jax.experimental import pallas as pl from jax.experimental.pallas import tpu as pltpu import jax.numpy as jnp

    DEFAULT_MASK_VALUE = -0.7 * float(jnp.finfo(jnp.dtype("float32")).max) NUM_L▌

  3. User

    You will recieve two files, one of them is a modification of the other, please note all the changes.File 1:

    """Flash Attention TPU kernel.""" from future import annotations

    import dataclasses import functools from typing import Any, NamedTuple

    import jax from jax import lax from jax.experimental import pallas as pl from jax.experimental.pallas import tpu as pltpu import jax.numpy as jnp

    DEFAULT_MASK_VALUE = -0.7 * float(jnp.finfo(jnp.dtype("float32")).max) NUM_LANES = 128 NUM_SUBLANES = 8

    class SegmentIds(NamedTuple): """SegmentIds for Q and KV sequences.

    SegmentIds are used to generate segment mask, which prevents attention between different segments in the input sequence. Each array is a list of ids (integers). Only the token with the same id can attend to each other.

    Attributes: q: segment ids along the Q sequence. kv: segment ids along the KV sequence. """

    q: jax.Array # [batch_size, q_seq_len] kv: jax.Array # [batch_size, kv_seq_len]

    @dataclasses.dataclass(frozen=True) class BlockSizes: """Tile sizes parameterizing FlashAttention kernels.

    Those parameters have negligible effect on numerics, but affect performance greatly. """ block_q: int block_k_major: int block_k: int block_b: int

    block_q_major_dkv: int | None = None block_k_major_dkv: int | None = None block_k_dkv: int | None = None block_q_dkv: int | None = None

    block_k_major_dq: int | None = None block_k_dq: int | None = None block_q_dq: int | None = None

    def post_init(self): def verify_major_minor(prefix, suffix, major, minor): if minor > major: raise ValueError( f"{prefix}{suffix}={minor} should be smaller than" f" {prefix}_major{suffix}={major}" ) if major % minor != 0: raise ValueError( f"{prefix}{suffix}={minor} should divide" f" {prefix}_major{suffix}={major}" )

    verify_major_minor("block_k", "", self.block_k_major, self.block_k)
    if self.block_q_major_dkv is not None and self.block_q_dkv is not None:
      verify_major_minor(
          "block_q", "_dkv", self.block_q_major_dkv, self.block_q_dkv
      )
    if self.block_k_major_dkv is not None and self.block_k_dkv is not None:
      verify_major_minor(
          "block_k", "_dkv", self.block_k_major_dkv, self.block_k_dkv
      )
    if self.block_k_major_dq is not None and self.block_k_dq is not None:
      verify_major_minor(
          "block_k", "_dq", self.block_k_major_dq, self.block_k_dq
      )
    

    @property def has_backward_blocks(self) -> bool: backward_blocks = ( self.block_q_major_dkv, self.block_k_major_dkv, self.block_q_dkv, self.block_k_dkv, self.block_k_major_dq, self.block_k_dq, self.block_q_dq, ) return all(b is not None for b in backward_blocks)

    @classmethod def get_default(cls, batch_size, num_heads, q_seq_len, kv_len, d_model): # TODO(apaszke,sharadmv): Select better parameters based on a heuristic. del batch_size, num_heads, q_seq_len, kv_len, d_model # Unused. return BlockSizes( block_q=128, block_k_major=128, block_k=128, block_b=1, block_q_major_dkv=128, block_k_major_dkv=128, block_k_dkv=128, block_q_dkv=128, block_k_major_dq=128, block_k_dq=128, block_q_dq=128, )

    @functools.partial( jax.jit, static_argnames=[ "causal", "sm_scale", "block_sizes", "debug", ], ) def flash_attention( q, # [batch_size, num_heads, q_seq_len, d_model] k, # [batch_size, num_heads, kv_seq_len, d_model] v, # [batch_size, num_heads, kv_seq_len, d_model] ab=None, # [batch_size, num_heads, q_seq_len, kv_seq_len] segment_ids=None, # q of [batch_size, q_seq_len] and kv of [batch_size, kv_seq_len] *, causal: bool = False, sm_scale: float = 1.0, block_sizes: BlockSizes | None = None, debug: bool = False, ): batch_size, num_heads, q_seq_len, d_model = q.shape batch_size_k, num_heads_k, kv_seq_len, d_model_k = k.shape batch_size_v, num_heads_v, kv_seq_len_v, d_model_v = v.shape if batch_size != batch_size_k or batch_size != batch_size_v: raise ValueError( f"Batch size mismatch: got {batch_size}, {batch_size_k} and" f" {batch_size_v} (for q, k, v respectively)" ) if num_heads != num_heads_k or num_heads != num_heads_v: raise ValueError( f"Head count mismatch: got {num_heads}, {num_heads_k}," f" {num_heads_v} (for q, k, v respectively)" ) if d_model != d_model_k: raise ValueError( f"Model dimension mismatch: got {d_model} and {d_model_k} (for q and k" " respectively)" ) if d_model != d_model_v: raise NotImplementedError( "V model dimension unequal to KV model dimension unsupported" ) if kv_seq_len != kv_seq_len_v: raise ValueError( f"KV sequence length mismatch: got {kv_seq_len} and {kv_seq_len_v}" ) if ab is not None: if ab.shape != (batch_size, num_heads, q_seq_len, kv_seq_len): raise ValueError( f"Attention bias shape mismatch: expected ({batch_size=}," f" {num_heads=}, {q_seq_len=}, {kv_seq_len=}), got {ab.shape}" ) if segment_ids is not None: if segment_ids.q.shape != (batch_size, q_seq_len): raise ValueError( f"Q segment ids shape mismatch: expected ({batch_size=}," f" {q_seq_len=},), got {segment_ids.q.shape}" ) if segment_ids.kv.shape != (batch_size, kv_seq_len): raise ValueError( f"KV segment ids shape mismatch: expected ({batch_size=}," f" {kv_seq_len=},), got {segment_ids.kv.shape}" ) if block_sizes is None: block_sizes = BlockSizes.get_default( batch_size, num_heads, q_seq_len, kv_seq_len, d_model ) return _flash_attention( q, k, v, ab, segment_ids, False, causal, sm_scale, block_sizes, debug )

    @functools.partial(jax.custom_vjp, nondiff_argnums=range(5, 10)) def _flash_attention( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_sizes, debug, ): return _flash_attention_impl( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_sizes.block_b, block_sizes.block_q, block_sizes.block_k_major, block_sizes.block_k, debug, )

    def _flash_attention_fwd( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_sizes, debug, ): if save_residuals: raise NotImplementedError("Higher-order AD not supported") o, l, m = _flash_attention( q, k, v, ab, segment_ids, True, causal, sm_scale, block_sizes, debug ) return o, (q, k, v, ab, segment_ids, o, l, m)

    def _flash_attention_bwd( save_residuals: bool, causal: bool, sm_scale: float, block_sizes: BlockSizes, debug: bool, residuals, do, ): """VJP rule for FlashAttention.""" if save_residuals: raise NotImplementedError("Higher-order AD not supported") (q, k, v, ab, segment_ids, o, l, m) = residuals if not block_sizes.has_backward_blocks: raise ValueError( "Program is being differentiated, but not all backward blocks are" " specified" )

    di = jnp.sum( o.astype(jnp.float32) * do.astype(jnp.float32), axis=-1 ) # [batch_size, num_heads, q_seq_len]

    dk, dv = _flash_attention_bwd_dkv( q, k, v, ab, segment_ids, l, m, do, di, block_q_major=block_sizes.block_q_major_dkv, block_k_major=block_sizes.block_k_major_dkv, block_k=block_sizes.block_k_dkv, block_q=block_sizes.block_q_dkv, sm_scale=sm_scale, causal=causal, mask_value=DEFAULT_MASK_VALUE, debug=debug, )

    dq, ds = _flash_attention_bwd_dq( q, k, v, ab, segment_ids, l, m, do, di, block_q_major=block_sizes.block_q_dq, block_k_major=block_sizes.block_k_major_dq, block_k=block_sizes.block_k_dq, sm_scale=sm_scale, causal=causal, mask_value=DEFAULT_MASK_VALUE, debug=debug, ) return dq, dk, dv, ds, None

    _flash_attention.defvjp(fwd=_flash_attention_fwd, bwd=_flash_attention_bwd)

    MIN_BLOCK_SIZE = 128 TRANS_B_DIM_NUMBERS = (((1,), (1,)), ((), ()))

    def below_or_on_diag(r, r_blk_size, c, c_blk_size):

    A block is considered below or on diagonal as long as the bottom left

    corner of the block is below or on diagonal.

    return ((r + 1) * r_blk_size - 1) > (c * c_blk_size)

    def _flash_attention_kernel(q_tile_ref, *args, **kwargs): block_b = q_tile_ref.shape[0]

    If we're not going to tile the softmax, then we can avoid a bunch of VPU ops.

    if kwargs["block_k"] == kwargs["kv_seq_len"]: kernel = _flash_attention_kernel_single_batch_single_step else: kernel = _flash_attention_kernel_single_batch for batch_idx in range(block_b): kernel((batch_idx, 0), q_tile_ref, *args, **kwargs)

    def _flash_attention_kernel_single_batch( batch_idx: tuple[int, ...], q_tile_ref, k_tile_ref, v_tile_ref, ab_tile_ref, q_segment_ids_tile_ref, kv_segment_ids_tile_ref, # Input arrays o_tile_ref, # Output arrays l_ref, m_ref, m_scratch_ref, l_scratch_ref, acc_scratch_ref, *, causal, sm_scale, block_k, kv_seq_len, mask_value, ): block_k_major = k_tile_ref.shape[2] block_q = q_tile_ref.shape[2] head_dim = q_tile_ref.shape[-1]

    kv_seq_idx = pl.program_id(3) @pl.when(kv_seq_idx == 0) def start_new_sequence(): m_scratch_ref[batch_idx] = jnp.full( m_scratch_ref.shape[2:], -jnp.inf, jnp.float32 ) l_scratch_ref[batch_idx] = jnp.zeros(l_scratch_ref.shape[2:], jnp.float32) acc_scratch_ref[batch_idx] = jnp.zeros( acc_scratch_ref.shape[2:], jnp.float32 )

    q_seq_idx = pl.program_id(2) if causal: should_run = below_or_on_diag(q_seq_idx, block_q, kv_seq_idx, block_k_major) else: should_run = True

    @pl.when(should_run) def run(): @functools.partial( lax.fori_loop, 0, block_k_major // block_k, init_val=None, unroll=True ) def body(i, _): m_prev = m_scratch_ref[batch_idx] l_prev = l_scratch_ref[batch_idx] q = q_tile_ref[batch_idx] # [block_q, head_dim] start_k = i * block_k k = pl.load( k_tile_ref, (*batch_idx, pl.dslice(start_k, block_k), slice(None)) ) # [block_k, head_dim]

      s = jax.lax.dot_general(
          q, k, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32
      )  # [block_q, block_k]
    
      # Add attention bias if needed.
      # TODO(tanburn) Should the attention bias be added before or after
      # multiplication by sm_scale?
      if ab_tile_ref is not None:
        ab = pl.load(
            ab_tile_ref,
            (*batch_idx, pl.dslice(None), pl.dslice(start_k, block_k))
        ).astype(jnp.float32)
        s += ab
    
      if sm_scale != 1.0:
        s *= sm_scale
    
      mask = None
      if q_segment_ids_tile_ref is not None:
        repeats, rem = divmod(block_k, NUM_LANES)
        if rem:
          raise NotImplementedError(
              f"kv block size must be a multiple of {NUM_LANES}"
          )
        q_segment_ids = pltpu.repeat(
            q_segment_ids_tile_ref[batch_idx[0]], repeats, axis=1
        )  # [block_q, block_k].
        kv_segment_ids = pl.load(
            kv_segment_ids_tile_ref,
            (batch_idx[0], pl.dslice(1), pl.dslice(start_k, block_k)),
        )  # [1, block_k].
        mask = jnp.equal(q_segment_ids, kv_segment_ids).astype(jnp.bool_)
    
      if causal:
        mask_shape = (block_q, block_k)
        row_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 0)
        row_ids += q_seq_idx * block_q
        col_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 1)
        col_ids += kv_seq_idx * block_k_major + start_k
        causal_mask = col_ids <= row_ids
        mask = (
            causal_mask if mask is None else jnp.logical_and(mask, causal_mask)
        )
    
      s = s if mask is None else s + jnp.where(mask, 0.0, mask_value)
    
      m_curr = jnp.max(s, axis=1)[:, None]  # Row max, shape [block_q, 1].
      m_next = jnp.maximum(m_prev, m_curr)  # Shape [block_q, 128].
    
      block_k_repeats, rem = divmod(block_k, MIN_BLOCK_SIZE)
      if rem:
        raise NotImplementedError(
            f"{block_k=} should be a multiple of {MIN_BLOCK_SIZE}"
        )
      p = jnp.exp(s - pltpu.repeat(m_next, block_k_repeats, 1))
    
      alpha = jnp.exp(m_prev - m_next)  # Shape [block_q, 128].
    
      l_corr = alpha * l_prev
    
      l_next = jnp.sum(p, axis=1)[:, None] + l_corr  # Shape [block_q, 128]
    
      head_dim_repeats, rem = divmod(head_dim, MIN_BLOCK_SIZE)
      l_broadcast = lambda l: pltpu.repeat(l, head_dim_repeats, 1)
      if rem:
        if head_dim_repeats == 0:
          l_broadcast = lambda l: l[:, :head_dim]
        else:
          raise NotImplementedError(
              f"{head_dim=} should be a multiple of {MIN_BLOCK_SIZE} if larger"
          )
      l_scratch_ref[batch_idx] = l_next
      m_scratch_ref[batch_idx] = m_next
    
      l_next_inv_safe = jnp.where(l_next == 0.0, 1.0, 1.0 / l_next)
      acc_scratch_ref[batch_idx] *= l_broadcast(l_corr * l_next_inv_safe)
      v = pl.load(
          v_tile_ref, (*batch_idx, pl.dslice(start_k, block_k), slice(None))
      )
      o_curr = jax.lax.dot(
          p.astype(v.dtype), v, preferred_element_type=jnp.float32
      )
      acc_scratch_ref[batch_idx] += o_curr * l_broadcast(l_next_inv_safe)
    

    @pl.when(kv_seq_idx == (kv_seq_len // block_k_major) - 1) def store_output(): o_tile_ref[batch_idx] = acc_scratch_ref[batch_idx].astype(o_tile_ref.dtype) if l_ref is not None: l_ref[batch_idx] = l_scratch_ref[batch_idx].astype(l_ref.dtype) if m_ref is not None: m_ref[batch_idx] = m_scratch_ref[batch_idx].astype(m_ref.dtype)

    def _flash_attention_kernel_single_batch_single_step( batch_idx: tuple[int, ...], q_tile_ref, k_tile_ref, v_tile_ref, ab_tile_ref, q_segment_ids_tile_ref, kv_segment_ids_tile_ref, # Input arrays o_tile_ref, # Output arrays l_ref: Any | None = None, m_ref: Any | None = None, *, causal, sm_scale, block_k, kv_seq_len, mask_value, ): block_k_major = k_tile_ref.shape[2] block_q = q_tile_ref.shape[2]

    assert kv_seq_len == block_k_major == block_k

    q = q_tile_ref[batch_idx] # [block_q, head_dim] k = k_tile_ref[batch_idx] # [block_k, head_dim] s = jax.lax.dot_general( q, k, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32 ) # [block_q, block_k]

    if ab_tile_ref is not None: s += ab_tile_ref[batch_idx].astype(jnp.float32) if sm_scale != 1.0: s *= sm_scale

    mask = None if q_segment_ids_tile_ref is not None: repeats, rem = divmod(block_k, NUM_LANES) if rem: raise NotImplementedError( f"kv block size must be a multiple of {NUM_LANES}" ) q_segment_ids = pl.load( q_segment_ids_tile_ref, (batch_idx[0],) ) # [block_q, NUM_LANES]. q_segment_ids = pltpu.repeat( q_segment_ids, repeats, axis=1 ) # [block_q, block_k]. kv_segment_ids = pl.load( kv_segment_ids_tile_ref, (batch_idx[0], pl.dslice(1)) ) # [1, block_k]. mask = jnp.equal(q_segment_ids, kv_segment_ids).astype(jnp.bool_)

    if causal: q_seq_idx = pl.program_id(2) mask_shape = (block_q, block_k) row_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 0) row_ids += q_seq_idx * block_q col_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 1) causal_mask = col_ids <= row_ids mask = causal_mask if mask is None else jnp.logical_and(mask, causal_mask) s = s if mask is None else s + jnp.where(mask, 0.0, mask_value)

    m = jnp.max(s, axis=1)[:, None] p = jnp.exp(s - m) l = jnp.sum(p, axis=1)[:, None] p /= l

    if m_ref is not None: m_ref[batch_idx] = lax.broadcast_in_dim(m, m_ref.shape[2:], range(2)) if l_ref is not None: l_ref[batch_idx] = lax.broadcast_in_dim(l, l_ref.shape[2:], range(2))

    v = v_tile_ref[batch_idx] o_tile_ref[batch_idx] = jax.lax.dot( p.astype(v.dtype), v, preferred_element_type=jnp.float32 ).astype(o_tile_ref.dtype)

    def _flash_attention_impl( q, k, v, ab, segment_ids, save_residuals, causal, sm_scale, block_b, block_q, block_k_major, block_k, debug, ): batch_size, num_heads, q_seq_len, head_dim = q.shape _, _, kv_seq_len, _ = k.shape _verify_block("block_q", "q_seq_len", block_q, q_seq_len, should_divide=False) _verify_block("block_k_major", "kv_seq_len", block_k_major, kv_seq_len) _verify_block("block_k", "kv_seq_len", block_k, kv_seq_len) _verify_block("block_b", "batch", block_b, batch_size, should_divide=False)

    TODO(apaszke): Tile over heads as well.

    grid = ( pl.cdiv(batch_size, block_b), num_heads, pl.cdiv(q_seq_len, block_q), kv_seq_len // block_k_major, )

    def q_index_map(batch_index, head_index, q_seq_index, _): return (batch_index, head_index, q_seq_index, 0)

    def kv_index_map(batch_index, head_index, q_seq_index, kv_seq_index): if causal: # If the kv block is skipped, prefetch the next valid kv block, i.e. the # 0th one to be used for the next block_q rows. next_kv_index = lax.select( below_or_on_diag(q_seq_index, block_q, kv_seq_index, block_k_major), kv_seq_index, 0, ) else: next_kv_index = kv_seq_index return (batch_index, head_index, next_kv_index, 0)

    def ab_index_map(batch_index, head_index, q_seq_index, kv_seq_index): if causal: should_run = below_or_on_diag( q_seq_index, block_q, kv_seq_index, block_k_major ) # If the ab block is skipped, prefetch the next valid ab block, i.e. the # 0th kv to be used for the next block_q rows. next_q_index = lax.select( should_run, q_seq_index, lax.select( q_seq_index == (q_seq_len // block_q) - 1, 0, q_seq_index + 1 ), ) next_kv_index = lax.select(should_run, kv_seq_index, 0) else: next_q_index = q_seq_index next_kv_index = kv_seq_index

    return (batch_index, head_index, next_q_index, next_kv_index)
    

    def o_index_map(batch_index, head_index, q_seq_index, _): return (batch_index, head_index, q_seq_index, 0)

    def lm_index_map(batch_index, head_index, q_seq_index, _): return (batch_index, head_index, q_seq_index, 0)

    kernel = functools.partial( _flash_attention_kernel, causal=causal, mask_value=DEFAULT_MASK_VALUE, sm_scale=sm_scale, block_k=block_k, kv_seq_len=kv_seq_len, ) out_shape = jax.ShapeDtypeStruct(shape=q.shape, dtype=q.dtype) out_shape = [out_shape] out_specs = [pl.BlockSpec((block_b, 1, block_q, head_dim), o_index_map)]

    if block_k != kv_seq_len: m_scratch = pltpu.VMEM((block_b, 1, block_q, MIN_BLOCK_SIZE), jnp.float32) l_scratch = pltpu.VMEM((block_b, 1, block_q, MIN_BLOCK_SIZE), jnp.float32) acc_scratch = pltpu.VMEM((block_b, 1, block_q, head_dim), jnp.float32) scratch_shapes = [m_scratch, l_scratch, acc_scratch] else: scratch_shapes = []

    if save_residuals: out_specs = [ *out_specs, pl.BlockSpec((block_b, 1, block_q, MIN_BLOCK_SIZE), lm_index_map), pl.BlockSpec((block_b, 1, block_q, MIN_BLOCK_SIZE), lm_index_map), ] l = jax.ShapeDtypeStruct( (batch_size, num_heads, q_seq_len, MIN_BLOCK_SIZE), dtype=jnp.float32 ) m = jax.ShapeDtypeStruct( (batch_size, num_heads, q_seq_len, MIN_BLOCK_SIZE), dtype=jnp.float32 ) out_shape = (*out_shape, l, m) else: out_specs = [*out_specs, None, None] out_shape = (*out_shape, None, None)

    ab_block_spec = ( pl.BlockSpec((block_b, 1, block_q, block_k_major), ab_index_map) if ab is not None else None)

    q_segment_ids_spec = kv_segment_ids_spec = None q_segment_ids = kv_segment_ids = None if segment_ids is not None:

    def q_segment_ids_index_map(batch_index, head_index, q_seq_index, _):
      del head_index
      return (batch_index, q_seq_index, 0)
    
    def kv_segment_ids_index_map(
        batch_index, head_index, q_seq_index, kv_seq_index
    ):
      del head_index
      if causal:
        next_kv_index = lax.select(
            below_or_on_diag(q_seq_index, block_q, kv_seq_index, block_k_major),
            kv_seq_index,
            0,
        )
      else:
        next_kv_index = kv_seq_index
      return (batch_index, 0, next_kv_index)
    
    q_segment_ids_spec = pl.BlockSpec(
        (block_b, block_q, NUM_LANES), q_segment_ids_index_map
    )
    kv_segment_ids_spec = pl.BlockSpec(
        (block_b, NUM_SUBLANES, block_k_major), kv_segment_ids_index_map
    )
    
    q_segment_ids = jax.lax.broadcast_in_dim(
        segment_ids.q,
        (batch_size, q_seq_len, NUM_LANES),
        (
            0,
            1,
        ),
    )
    kv_segment_ids = jax.lax.broadcast_in_dim(
        segment_ids.kv,
        (batch_size, NUM_SUBLANES, kv_seq_len),
        (
            0,
            2,
        ),
    )
    

    in_specs = [ pl.BlockSpec((block_b, 1, block_q, head_dim), q_index_map), pl.BlockSpec((block_b, 1, block_k_major, head_dim), kv_index_map), pl.BlockSpec((block_b, 1, block_k_major, head_dim), kv_index_map), ab_block_spec, q_segment_ids_spec, kv_segment_ids_spec, ]

    o, *aux = pl.pallas_call( kernel, grid_spec=pltpu.PrefetchScalarGridSpec( num_scalar_prefetch=0, grid=grid, in_specs=in_specs, out_specs=out_specs, scratch_shapes=scratch_shapes, ), out_shape=out_shape, debug=debug, compiler_params=dict( mosaic=dict( dimension_semantics=( "parallel", "parallel", "parallel", "arbitrary", ) ) ), )(q, k, v, ab, q_segment_ids, kv_segment_ids) if save_residuals: l, m = (v[..., 0] for v in aux[-2:]) return (o, l, m) else: return o

    def _flash_attention_dkv_kernel( q_tile_ref, k_tile_ref, v_tile_ref, ab_tile_ref, q_segment_ids_tile_ref, kv_segment_ids_tile_ref, l_tile_ref, m_tile_ref, do_tile_ref, di_tile_ref, dk_tile_ref, dv_tile_ref, dk_scratch_ref, dv_scratch_ref, *, sm_scale: float, causal: bool, mask_value: float, q_seq_len: int, block_q: int, block_k: int, ): _, _, block_q_major, _ = q_tile_ref.shape _, _, block_k_major, _ = k_tile_ref.shape

    q_seq_index = pl.program_id(axis=3) kv_seq_index = pl.program_id(axis=2)

    @pl.when(q_seq_index == 0) def start_new_sequence(): dk_scratch_ref[:, :] = jnp.zeros(dk_scratch_ref.shape, dk_scratch_ref.dtype) dv_scratch_ref[:, :] = jnp.zeros(dv_scratch_ref.shape, dv_scratch_ref.dtype)

    def q_body(j, _): start_q = j * block_q def k_body(i, _): start_k = i * block_k k = pl.load(k_tile_ref, (0, 0, pl.ds(start_k, block_k), slice(None))) v = pl.load(v_tile_ref, (0, 0, pl.ds(start_k, block_k), slice(None))) q = pl.load(q_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, head_dim] l = pl.load(l_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, 128] m = pl.load(m_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, 128] do = pl.load(do_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ) # [block_q, 128] di = pl.load(di_tile_ref, (0, 0, pl.ds(start_q, block_q), slice(None)) ).astype(jnp.float32) # [block_q, 128]

      capped_logits = lax.dot_general(
          q, k, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32
      )  # [block_q_major, block_k]
    
      if ab_tile_ref is not None:
        ab = pl.load(
            ab_tile_ref,
            (
                0,
                0,
                pl.dslice(j * block_q, block_q),
                pl.dslice(i * block_k, block_k),
            ),
        ).astype(jnp.float32)
        capped_logits += ab
    
      if sm_scale != 1.0:
        capped_logits *= sm_scale
    
      mask = None
      if q_segment_ids_tile_ref is not None:
        repeats, rem = divmod(block_k, NUM_LANES)
        if rem:
          raise NotImplementedError(
          )
        q_segment_ids = pl.load(
            q_segment_ids_tile_ref, (0, pl.ds(start_q, block_q), slice(None))
        )  # [block_q, NUM_LANES].
        q_segment_ids = pltpu.repeat(
            q_segment_ids, repeats, axis=1
        )  # [block_q, block_k].
        kv_segment_ids = pl.load(
            kv_segment_ids_tile_ref, (slice(None), 0, pl.ds(start_k, block_k))
        )  # [1, block_k].
        mask = jnp.equal(q_segment_ids, kv_segment_ids).astype(jnp.bool_)
    
      if causal:
        mask_shape = (block_q, block_k)
        row_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 0)
        row_ids += q_seq_index * block_q_major + start_q
        col_ids = jax.lax.broadcasted_iota(jnp.int32, mask_shape, 1)
        col_ids += kv_seq_index * block_k_major + start_k
        causal_mask = col_ids <= row_ids
        mask = (
            causal_mask if mask is None else jnp.logical_and(mask, causal_mask)
        )
    
      capped_logits = (
          capped_logits
          if mask is None
          else capped_logits + jnp.where(mask, 0.0, mask_value)
      )
    
      p = jnp.exp(
          capped_logits - pltpu.repeat(m, block_k // MIN_BLOCK_SIZE, axis=1)
      )
      p = p * pltpu.repeat(
          1 / l, block_k // MIN_BLOCK_SIZE, axis=1
      )  # [block_q_major, block_k_major]
      dv = lax.dot(p.T.astype(do.dtype), do, preferred_element_type=jnp.float32)
      pl.store(dv_scratch_ref, (pl.ds(start_k, block_k), slice(None)),
               pl.load(dv_scratch_ref, (pl.ds(start_k, block_k), slice(None)))
               + dv.astype(dv_scratch_ref.dtype))
    
      # di: [block_q, 128]
      # do: [block_q, head_dim]
      # v: [block_k_major, head_dim]
      dp = lax.dot_general(
          do, v, TRANS_B_DIM_NUMBERS, preferred_element_type=jnp.float32
      )
      ds = (dp - pltpu.repeat(di, block_k // MIN_BLOCK_SIZE, axis=1)) * p
    
      if sm_scale != 1.0:
        ds = ds * sm_scale
    
      # ds: [block_q_major, block_k_major]
      # q: [block_q_major, head_dim]
      dk = lax.dot(ds.T.astype(do.dtype), q, preferred_element_type=jnp.float32)
      pl.store(dk_scratch_ref, (pl.ds(start_k, block_k), slice(None)),
               pl.load(dk_scratch_ref, (pl.ds(start_k, block_k), slice(None)))
               + dk.astype(dk_scratch_ref.dtype))
    lax.fori_loop(0, block_k_major // block_k, k_body, None, unroll=True)
    

    if causal: should_run = below_or_on_diag( q_seq_index, block_q_major, kv_seq_index, block_k_major ) else: should_run = True

    @pl.when(should_run) def run(): lax.fori_loop(0, block_q_major // block_q, q_body, None, unroll=True)

    @pl.when(q_seq_index == q_seq_len // block_q_major - 1) def end_of_q_sequence(): dv_tile_ref[0, 0, :, :] = dv_scratch_ref[...].astype(dv_tile_ref.dtype) dk_tile_ref[0, 0, :, :] = dk_scratch_ref[...].astype(dk_tile_ref.dtype)

    def _flash_attention_bwd_dkv( q, k, v, ab, segment_ids, l, m, do, di, *, block_q_major: int | None, block_q: int | None, block_k_major: int | None, block_k: int | None, sm_scale: float, causal: bool = False, mask_value: float = DEFAULT_MASK_VALUE, debug: bool = False, ): batch_size, num_heads, q_seq_len, head_dim = q.shape _, _, kv_seq_len, _ = k.shape _verify_block("block_q_major_dkv", "q_seq_len", block_q_major, q_seq_len) _verify_block("block_q_dkv", "q_seq_len", block_q, q_seq_len) _verify_block("block_k_major_dkv", "kv_seq_len", block_k_major, kv_seq_len) _verify_block("block_k_dkv", "kv_seq_len", block_k, kv_seq_len)

    Broadcast out scalar values

    m = jnp.broadcast_to(m[..., None], (*m.shape, MIN_BLOCK_SIZE)) l = jnp.broadcast_to(l[..., None], (*l.shape, MIN_BLOCK_SIZE))

    Preprocess contraction for bwd pass

    di = jnp.broadcast_to(di[..., None], (*di.shape, MIN_BLOCK_SIZE))

    kv index needs to be before q index since q index is the contractng

    dimension.

    grid = ( batch_size, num_heads, kv_seq_len // block_k_major, q_seq_len // block_q_major, )

    def qo_index_map(batch_index, head_index, kv_seq_index, q_seq_index): if causal: # If the q block is skipped, stay at the 0th q block. next_q_index = lax.select( below_or_on_diag( q_seq_index, block_q_major, kv_seq_index, block_k_major ), q_seq_index, 0, ) else: next_q_index = q_seq_index

    return (batch_index, head_index, next_q_index, 0)
    

    qo_spec = pl.BlockSpec((1, 1, block_q_major, head_dim), qo_index_map) assert qo_spec.block_shape is not None assert q.ndim == len(qo_spec.block_shape) do_spec = qo_spec assert do.ndim == len(qo_spec.block_shape)

    def kv_index_map(batch_index, head_index, kv_seq_index, _): return (batch_index, head_index, kv_seq_index, 0)

    kv_spec = pl.BlockSpec((1, 1, block_k_major, head_dim), kv_index_map) assert kv_spec.block_shape is not None assert k.ndim == len(kv_spec.block_shape) assert v.ndim == len(kv_spec.block_shape)

    def lm_index_map(batch_index, head_index, _, q_seq_index): return (batch_index, head_index, q_seq_index, 0)

    lm_spec = pl.BlockSpec((1, 1, block_q_major, MIN_BLOCK_SIZE), lm_index_

  4. Assistant

    File 2:

    """Flash Attention TPU kernel.""" from future import annotations

    import dataclasses import functools from typing import Any, NamedTuple

    import jax from jax import lax from jax.experimental import pallas as pl from jax.experimental.pallas import tpu as pltpu import jax.numpy as jnp

    DEFAULT_MASK_VALUE = -0.7 * float(jnp.finfo(jnp.dtype("float32")).max) NUM_LANES = 128▌

Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.