Source code for easydel.layers.attention_operator.modules.splash

# Copyright 2023 The EASYDEL Author @erfanzar (Erfan Zare Chavoshi).
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.


import functools
import typing as tp

import jax
from jax import Array
from jax import numpy as jnp
from jax import random as jr
from jax.experimental.pallas.ops.tpu.splash_attention import (
	BlockSizes,
	CausalMask,
	MultiHeadMask,
	SegmentIds,
	make_splash_mqa_single_device,
)
from jax.experimental.shard_map import shard_map
from jax.sharding import PartitionSpec as Ps

from .._attention_impl import (
	AttentionImpl,
	AttentionMetadata,
	AttentionOutput,
	AttentionRegistry,
)
from .vanilla import VanillaAttn


[docs]@AttentionRegistry.register class SplashAttn(AttentionImpl):
[docs] @classmethod def get_impl_name(cls) -> tp.Union[str, tp.Tuple[str]]: return "splash"
[docs] def get_impl_metadata(self) -> AttentionMetadata: return self.metadata
[docs] def forward_native(self, *args, **kwargs) -> AttentionOutput: raise NotImplementedError("`forward_native` not implemented!")
[docs] def forward_gpu(self, *args, **kwargs) -> AttentionOutput: raise NotImplementedError("`forward_gpu` not implemented!")
[docs] @jax.named_scope("easydel-splashimpl-tpu") def forward_tpu( self, q: Array, k: Array, v: Array, mask: tp.Optional[Array] = None, causal: bool = True, **ignore, ) -> AttentionOutput: query_lenght = q.shape[1] value_lenght = v.shape[1] if (query_lenght == 1) or not causal or ((query_lenght % 128) != 0): return VanillaAttn(self.metadata)( q=q, k=k, v=v, mask=mask, causal=causal, ) sm_scale = self.metadata.softmax_scale sm_scale = sm_scale if sm_scale is not None else q.shape[-1] ** -0.5 dtype = self.metadata.runtime_dtype runtime_type = self.get_runtime_type(q=q, BTHD=False) ( query_partition_spec, key_partition_spec, value_partition_spec, bias_partition_spec, mask_partition_spec, attention_partition_spec, ) = self.metadata.get_partition_specs(runtime_type, BTHD=False) if mask is not None and mask.shape[0] != q.shape[0]: num_reps_mask = q.shape[0] // mask.shape[0] mask = jnp.repeat(mask, num_reps_mask, 0) block_sizes = BlockSizes( block_q=min(self.metadata.blocksize_q, query_lenght), block_kv_compute=min(self.metadata.blocksize_k, value_lenght), block_kv=min(self.metadata.blocksize_k, value_lenght), block_q_dkv=min(self.metadata.blocksize_q, query_lenght), block_kv_dkv=min(self.metadata.blocksize_k, value_lenght), block_kv_dkv_compute=min(self.metadata.blocksize_k, value_lenght), block_q_dq=min(self.metadata.blocksize_q, query_lenght), block_kv_dq=min(self.metadata.blocksize_k, value_lenght), ) qkv_mask_partition_spec = Ps(query_partition_spec[0], query_partition_spec[2]) q_mask, kv_mask = [None] * 2 if mask is not None: q_mask, kv_mask = self._split_attention_mask(mask) q_mask, kv_mask = ( q_mask.astype("i4"), kv_mask.astype("i4"), ) # pallas dont support int1 or bool in shardmap idk why pi = [0, 1, 3] mpi = [0] # query_partition_spec is like PB,PH,PS,PD # v is like BSHD since it's not transposed yet tparallel = self.metadata.mesh.shape[query_partition_spec[1]] if (v.shape[2] % tparallel) == 0 and tparallel <= v.shape[2]: pi = [0, 3] # shard DP, FSDP and TP @functools.partial( shard_map, mesh=self.metadata.mesh, in_specs=( self.create_stable_sharding(query_partition_spec, pi, dep=q), self.create_stable_sharding(key_partition_spec, pi, dep=k), self.create_stable_sharding(value_partition_spec, pi, dep=v), self.create_stable_sharding(qkv_mask_partition_spec, mpi, dep=q_mask), self.create_stable_sharding(qkv_mask_partition_spec, mpi, dep=kv_mask), ), out_specs=self.create_stable_sharding(attention_partition_spec, pi), check_rep=False, ) def _wraped_flash_attn(q, k, v, q_mask, kv_mask): output_shape = q.shape[:-1] + (v.shape[-1],) num_reps = q.shape[1] // k.shape[1] q = q.reshape(q.shape[:-3] + (k.shape[-3], num_reps, q.shape[-2], q.shape[-1])) fn = jax.vmap( jax.vmap( make_splash_mqa_single_device( mask=MultiHeadMask( [CausalMask((q.shape[-2], k.shape[-2])) for _ in range(q.shape[-3])] ), block_sizes=block_sizes, ), in_axes=(0, 0, 0, None), ), in_axes=(0, 0, 0, 0), ) m = None if kv_mask is not None: m = SegmentIds(q_mask, kv_mask) return fn(q * sm_scale, k, v, m).reshape(output_shape) attn = _wraped_flash_attn( q.transpose(0, 2, 1, 3).astype(dtype), k.transpose(0, 2, 1, 3).astype(dtype), v.transpose(0, 2, 1, 3).astype(dtype), q_mask, kv_mask, ).transpose(0, 2, 1, 3) return AttentionOutput(attention_weights=None, attention_outputs=attn)
[docs] def forward_cpu(self, *args, **kwargs) -> AttentionOutput: raise NotImplementedError("`forward_cpu` not implemented!")
[docs] def forward_cuda(self, *args, **kwargs) -> AttentionOutput: raise NotImplementedError("`forward_cuda` not implemented!")
[docs] def forward_rocm(self, *args, **kwargs) -> AttentionOutput: raise NotImplementedError("`forward_rocm` not implemented!")
def __call__( self, q: Array, k: Array, v: Array, mask: tp.Optional[Array] = None, causal: bool = True, **ignore, ) -> AttentionOutput: return super().__call__(q=q, k=k, v=v, mask=mask, causal=causal)
if __name__ == "__main__": from easydel.infra import EasyDeLBaseConfig test_cases = [ # (batch_size, q_seq_len, k_seq_len, q_heads, k_heads) (1, 2048, 2048, 32, 4), (2, 2**13, 2**13, 32, 8), (4, 2**14, 2**14, 16, 8), (4, 2**13, 2**14, 16, 4), ] metadata = AttentionMetadata( runtime_dtype=jnp.bfloat16, base_config=EasyDeLBaseConfig(axis_dims=(1, 1, 1, -1)), ) splash_attn = SplashAttn(metadata) vanilla_attn = VanillaAttn(metadata) for idx, (b, qs, ks, qh, kh) in enumerate(test_cases): d, vd = 128, 128 print( f"Running test case {idx + 1}/{len(test_cases)}: " f"b={b}, qs={qs}, ks={ks}, qh={qh}, kh={kh}, d={d}, vd={vd}" ) key_q, key_k, key_v = jr.split(jr.PRNGKey(0), 3) q = jr.normal(key_q, (b, qs, qh, d), dtype=jnp.float32) k = jr.normal(key_k, (b, ks, kh, d), dtype=jnp.float32) v = jr.normal(key_v, (b, ks, kh, vd), dtype=jnp.float32) mask = SplashAttn._create_causal_mask(max(qs, ks))[-qs:, :ks] mask = jnp.broadcast_to(mask, (b, 1, qs, ks)) splash_out = splash_attn(q=q, k=k, v=v, mask=None).attention_outputs vanilla_out = vanilla_attn(q=q, k=k, v=v, mask=None).attention_outputs is_close = jnp.allclose(splash_out, vanilla_out, atol=0.125) max_diff = jnp.max(jnp.abs(splash_out - vanilla_out)) print(f"Test case {idx + 1} result: {'PASS' if is_close else 'FAIL'}") print(f"Maximum absolute difference: {max_diff}\n")