tpu-inference 0.12.0.dev20251222__py3-none-any.whl

tpu_inference/layers/jax/attention/deepseek_v3_attention.py

@@ -0,0 +1,547 @@
+# Copyright 2025 Google LLC
+#
+# 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
+#
+#     http://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 math
+from dataclasses import InitVar, dataclass
+from typing import Any, Tuple
+
+import jax
+import jax.numpy as jnp
+from flax import nnx
+from flax.typing import Sharding
+from jax.sharding import Mesh
+from jax.sharding import PartitionSpec as P
+
+from tpu_inference import utils
+from tpu_inference.kernels.mla.v1.kernel import mla_ragged_paged_attention
+from tpu_inference.kernels.ragged_paged_attention.v3.kernel import \
+    ragged_paged_attention
+from tpu_inference.kernels.ragged_paged_attention.v3.tuned_block_sizes import \
+    get_tuned_block_sizes
+from tpu_inference.layers.common.attention_metadata import AttentionMetadata
+from tpu_inference.layers.common.sharding import ShardingAxisName
+from tpu_inference.layers.jax.base import create_param
+from tpu_inference.layers.jax.layers import RMSNorm
+from tpu_inference.layers.jax.rope import DeepseekScalingRotaryEmbedding
+
+KVCache = Tuple[jax.Array, jax.Array]
+
+
+# TODO (wenxindongwork): Add MLA KV cache implementation. For now, cache complete KV vectors.
+@dataclass(kw_only=True)
+class MLA(nnx.Module):
+    """An implementation of Multi-Head Latent Attention as
+    described in the DeepSeek V3 paper.
+
+    Attributes:
+        mesh: The JAX device mesh for distributed computation.
+    """
+    hidden_size: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    head_dim: int
+    rope_theta: float
+    rope_scaling: dict[str, Any]
+    dtype: jnp.dtype
+    kv_cache_dtype: str
+    mesh: Mesh
+
+    q_lora_rank: int
+    kv_lora_rank: int
+    qk_nope_head_dim: int
+    qk_rope_head_dim: int
+    v_head_dim: int
+    rms_norm_eps: float
+
+    # Sharding attributes
+    rd_sharding: Sharding = ()
+    q_da_sharding: Sharding = ()
+    ap_sharding: Sharding = ()
+    anh_sharding: Sharding = ()
+    kv_da_sharding: Sharding = ()
+
+    activation_attention_td: Sharding = ()
+    activation_q_td: Sharding = ()
+    query_tnh: P = P()
+    keyvalue_skh: P = P()
+
+    attn_o_tnh: P = P()
+    activation_attention_out_td: Sharding = ()
+
+    random_init: bool = False
+    attention_chunk_size: int | None = None
+    rope_input_ordering: str = "split"
+    quant: Any | None = None
+    rope_mscale_all_dim: float = 1.0
+    use_mla_kernel: bool = False
+
+    rngs: InitVar[nnx.Rngs]
+
+    _q_scale: float = 1
+    _k_scale: float = 1
+    _v_scale: float = 1
+
+    def __post_init__(self, rngs: nnx.Rngs):
+        self.N = self.num_attention_heads
+        self.K = self.num_key_value_heads
+        self.D = self.hidden_size
+        self.qk_head_dim = self.qk_nope_head_dim + self.qk_rope_head_dim
+
+        if not self.use_mla_kernel:
+            assert self.N == self.K, "N and K must be equal for MLA"
+
+        if self.rope_scaling["factor"] <= 1.0:
+            yarn_mscale = 1.0
+        else:
+            yarn_mscale = 0.1 * self.rope_mscale_all_dim * math.log(
+                self.rope_scaling["factor"]) + 1.0
+        self.scale = self.qk_head_dim**-0.5 * yarn_mscale**2
+
+        self.rope = DeepseekScalingRotaryEmbedding(
+            rotary_dim=self.qk_rope_head_dim,
+            rope_theta=self.rope_theta,
+            original_max_position_embeddings=self.
+            rope_scaling["original_max_position_embeddings"],
+            scaling_factor=self.rope_scaling["factor"],
+            dtype=self.dtype,
+            beta_fast=self.rope_scaling["beta_fast"],
+            beta_slow=self.rope_scaling["beta_slow"],
+            mscale_value=self.rope_scaling["mscale"],
+            mscale_all_dim=self.rope_scaling["mscale_all_dim"],
+        )
+
+        # Initializes the weight kernels
+        self.kernel_q_down_proj_DA = create_param(rngs,
+                                                  (self.D, self.q_lora_rank),
+                                                  self.q_da_sharding,
+                                                  self.dtype,
+                                                  random_init=self.random_init)
+        self.kernel_q_up_proj_AP = create_param(
+            rngs,
+            (self.q_lora_rank, self.N * self.qk_head_dim),
+            self.ap_sharding,
+            self.dtype,
+            random_init=self.random_init,
+        )
+        self.kernel_kv_down_proj_DA = create_param(
+            rngs,
+            (self.D, self.kv_lora_rank + self.qk_rope_head_dim),
+            self.kv_da_sharding,
+            self.dtype,
+            random_init=self.random_init,
+        )
+        # NOTE (jacobplatin): we are keeping these variables as 3D because
+        # we would need to reshape them before the below projection,
+        # which caused issues as Qwix wasn't quantizing it correctly
+        # on the abstract pass
+        if self.use_mla_kernel:
+            self.kernel_k_up_proj_ANH = create_param(
+                rngs,
+                (self.kv_lora_rank, self.N, self.qk_nope_head_dim),
+                self.anh_sharding,
+                self.dtype,
+                random_init=self.random_init,
+            )
+            self.kernel_v_up_proj_ANH = create_param(
+                rngs,
+                (self.kv_lora_rank, self.N, self.v_head_dim),
+                self.anh_sharding,
+                self.dtype,
+                random_init=self.random_init,
+            )
+        else:
+            self.kernel_kv_up_proj_AL = create_param(
+                rngs,
+                (self.kv_lora_rank, self.N *
+                 (self.qk_nope_head_dim + self.v_head_dim)),
+                self.
+                ap_sharding,  # NOTE: we use the same sharding for kv_up_proj_AL and kernel_q_up_proj_AP
+                self.dtype,
+                random_init=self.random_init,
+            )
+        self.kernel_o_proj_RD = create_param(
+            rngs, (self.N * self.v_head_dim, self.D),
+            self.rd_sharding,
+            self.dtype,
+            random_init=self.random_init)
+        self.q_rms_norm = RMSNorm(
+            dims=self.q_lora_rank,
+            epsilon=self.rms_norm_eps,
+            with_scale=True,
+            dtype=self.dtype,
+            random_init=self.random_init,
+            rngs=rngs,
+        )
+
+        self.kv_rms_norm = RMSNorm(
+            dims=self.kv_lora_rank,
+            random_init=self.random_init,
+            epsilon=self.rms_norm_eps,
+            with_scale=True,
+            dtype=self.dtype,
+            rngs=rngs,
+        )
+
+        self.kv_cache_quantized_dtype = None
+        if self.kv_cache_dtype != "auto":
+            self.kv_cache_quantized_dtype = utils.get_jax_dtype_from_str_dtype(
+                self.kv_cache_dtype)
+
+    def __call__(self,
+                 x,
+                 is_prefill,
+                 kv_cache: KVCache,
+                 attention_metadata: AttentionMetadata,
+                 use_attention_rope: bool = True):
+        """Performs the forward pass of the attention module.
+
+        Args:
+            x: The input tensor of shape `(batch_size, seq_len, d_model)`.
+            is_prefill: Whether the operation mode is prefill (otherwise it is generate).
+            kv_cache: The key-value cache for storing past attention states.
+            attention_metadata: Metadata for attention, such as input positions.
+
+        Returns:
+            A tuple containing:
+                - The updated KV cache.
+                - The attention output tensor of shape
+                  `(batch_size, seq_len, d_model)`.
+        """
+        md = attention_metadata
+        x = jnp.asarray(x, self.dtype)
+        x_SD = nnx.with_sharding_constraint(x, self.activation_attention_td)
+        x_q_TD = nnx.with_sharding_constraint(x, self.activation_q_td)
+
+        with jax.named_scope("q_proj"):
+            # Query down projection.
+            q_TA = jnp.einsum("TD,DA -> TA", x_q_TD,
+                              self.kernel_q_down_proj_DA.value)
+            q_TA = self.q_rms_norm(q_TA)
+            # Query up projection, then reshape to TNH.
+            q_TP = jnp.einsum("TA,AP -> TP", q_TA,
+                              self.kernel_q_up_proj_AP.value)
+            q_TNH = q_TP.reshape(q_TA.shape[0], self.N, self.qk_head_dim)
+            # Split the query into nope and rope.
+            q_nope_TNH = q_TNH[..., :self.qk_nope_head_dim]
+            q_rope_TNH = q_TNH[..., self.qk_nope_head_dim:]
+            q_rope_TNH = self.rope.apply_rope(md.input_positions, q_rope_TNH)
+            if self.use_mla_kernel:
+                # Absorb the k up-projection matrix into q
+                q_TNA = jnp.einsum("TNH,ANH -> TNA", q_nope_TNH,
+                                   self.kernel_k_up_proj_ANH.value)
+                q_TNA = nnx.with_sharding_constraint(q_TNA, self.query_tnh)
+            else:
+                # Concatenate the nope and rope queries.
+                q_TNH = jnp.concatenate([q_nope_TNH, q_rope_TNH], axis=-1)
+                # Multiply the query by scaling factor
+                q_TNH = nnx.with_sharding_constraint(q_TNH, self.query_tnh)
+
+        with jax.named_scope("kv_proj"):
+            # KV down projection.
+            kv_SA = jnp.einsum("SD,DA -> SA", x_SD,
+                               self.kernel_kv_down_proj_DA.value)
+            # Split the key and value into latent kv vector and k rope vector.
+            k_rope_SH = kv_SA[..., self.kv_lora_rank:]
+            # Reshape k_rope_BSH to include head dimension for RoPE application
+            k_rope_SNH = k_rope_SH[..., None, :]
+            k_rope_SNH = self.rope.apply_rope(md.input_positions, k_rope_SNH)
+            assert k_rope_SNH.shape[1] == 1
+            k_rope_SH = k_rope_SNH[:, 0, :]
+
+            kv_SA = kv_SA[..., :self.kv_lora_rank]
+            kv_SA = self.kv_rms_norm(kv_SA)
+            kv_SA = nnx.with_sharding_constraint(kv_SA, self.keyvalue_skh)
+
+            if not self.use_mla_kernel:
+                k_rope_SNH = jnp.broadcast_to(
+                    k_rope_SNH,
+                    (k_rope_SNH.shape[0], self.N, self.qk_rope_head_dim))
+                # KV up projection, then reshape to SN(Hk+Hv).
+                kv_SL = jnp.einsum("SA,AL -> SL", kv_SA,
+                                   self.kernel_kv_up_proj_AL.value)
+                kv_nope_SNH = kv_SL.reshape(
+                    kv_SA.shape[0], self.N,
+                    self.qk_nope_head_dim + self.v_head_dim)
+                # Split the latent kv vector into k nope vector and v vector.
+                k_nope_SNH = kv_nope_SNH[..., :self.qk_nope_head_dim]
+                v_SNH = kv_nope_SNH[..., self.qk_nope_head_dim:]
+                # Concatenate the key vector.
+                k_SNH = jnp.concatenate([k_nope_SNH, k_rope_SNH], axis=-1)
+                k_SNH = nnx.with_sharding_constraint(k_SNH, self.keyvalue_skh)
+                v_SNH = nnx.with_sharding_constraint(v_SNH, self.keyvalue_skh)
+
+        with jax.named_scope("attn_op"):
+            # TODO(wenxindongwork): K and V have different head dimension,
+            # which is not supported by the current kv cache implementation.
+            # For now we are padding the v dimension to match the k dimension.
+            # Furthermore, deepseekv3 k head dimension is 192, which is
+            # not supported by the current attention kernel, which expects
+            # q, k, v head dimension to be multiple of 128. For now, we will
+            # pad the q, k, v dimension to multiple of 128.
+            # We should update the MLA kv cache implementation in the future.
+            if not self.use_mla_kernel:  # MLA kernel handles padding
+                multiple_of_128 = ((self.qk_head_dim - 1) // 128 + 1) * 128
+                q_TNH = jnp.pad(q_TNH,
+                                ((0, 0), (0, 0),
+                                 (0, multiple_of_128 - self.qk_head_dim)))
+                k_SNH = jnp.pad(k_SNH,
+                                ((0, 0), (0, 0),
+                                 (0, multiple_of_128 - self.qk_head_dim)))
+                v_SNH = jnp.pad(v_SNH,
+                                ((0, 0), (0, 0),
+                                 (0, multiple_of_128 - self.v_head_dim)))
+
+            q_scale = k_scale = v_scale = None
+
+            # TODO(gpolovets): MLA does not currently support quantized KV!
+            if not self.use_mla_kernel:
+                if self.kv_cache_quantized_dtype:
+                    # TODO(kyuyeunk/jacobplatin): Enable w8a8 when VREG spill issue is resolved.
+                    k_scale = self._k_scale
+                    v_scale = self._v_scale
+                    k_SNH, v_SNH = utils.quantize_kv(
+                        k_SNH, v_SNH, self.kv_cache_quantized_dtype, k_scale,
+                        v_scale)
+
+                new_kv_cache, outputs_TNH = self.attention(
+                    is_prefill,
+                    kv_cache,
+                    q_TNH,
+                    k_SNH,
+                    v_SNH,
+                    attention_metadata,
+                    self.mesh,
+                    q_scale,
+                    k_scale,
+                    v_scale,
+                )
+                # TODO(wenxindongwork): For now, unpad the outputs_TNH to match the v_head_dim.
+                # We shall add the MLA kv cache implementation in the future.
+                outputs_TNH = outputs_TNH[..., :self.v_head_dim]
+
+            else:
+                new_kv_cache, outputs_TNA = self.mla_attention(
+                    kv_cache,
+                    q_TNA,
+                    q_rope_TNH,
+                    kv_SA,
+                    k_rope_SH,
+                    attention_metadata,
+                    self.mesh,
+                )
+                outputs_TNH = jnp.einsum("TNA,ANH -> TNH", outputs_TNA,
+                                         self.kernel_v_up_proj_ANH.value)
+
+            with jax.named_scope("o_proj"):
+                outputs_TNH = nnx.with_sharding_constraint(
+                    outputs_TNH, self.activation_attention_out_td)
+                outputs_TR = outputs_TNH.reshape(outputs_TNH.shape[0],
+                                                 self.N * self.v_head_dim)
+                o_TD = jnp.einsum("TR,RD -> TD", outputs_TR,
+                                  self.kernel_o_proj_RD.value)
+
+            return new_kv_cache, o_TD
+
+    def attention(
+        self,
+        is_prefill: bool,
+        kv_cache: KVCache,
+        q_TNH: jax.Array,
+        k_SKH: jax.Array,
+        v_SKH: jax.Array,
+        attention_metadata: AttentionMetadata,
+        mesh: Mesh,
+        q_scale: float | None = None,
+        k_scale: float | None = None,
+        v_scale: float | None = None,
+    ) -> Tuple[KVCache, jax.Array]:
+        """Performs scaled dot-product attention and updates the KV cache.
+
+        This function handles the core attention logic, which varies between
+        prefill and generation modes. In prefill, it computes self-attention
+        over the input sequence with a causal mask. In generation, it attends
+        to the full history of keys and values stored in the cache.
+
+        Args:
+            is_prefill: A boolean indicating if the mode is 'prefill'.
+            kv_cache: The key-value cache to be updated and used.
+            q_TNH: Query tensor of shape `(query_seq, num_attention_heads, head_dim)`.
+            k_SKH: Key tensor of shape `(kv_seq, num_key_value_heads, head_dim)`.
+            v_SKH: Value tensor of shape `(kv_seq, num_key_value_heads, head_dim)`.
+            attention_metadata: Metadata containing sequence lengths.
+            mesh: The JAX device mesh (unused in this specific function but
+                kept for potential future use or API consistency).
+            q_scale: Quantization scale for q.
+            k_scale: Quantization scale for k.
+            v_scale: Quantization scale for v.
+
+        Returns:
+            A tuple containing:
+                - The updated KV cache.
+                - The attention output tensor of shape
+                  `(seq, num_q_heads, head_dim)`.
+        """
+        md = attention_metadata
+        in_specs = (
+            self.query_tnh,  # q
+            self.keyvalue_skh,  # k
+            self.keyvalue_skh,  # v
+            P(None, None, "model"),  # kv_cache
+            P(),  # md.seq_lens: Replicated
+            P(),  # page_indices_flat: Replicated
+            P(),  # query_start_loc: Replicated
+            P(),  # distribution: Replicated
+        )
+        out_specs = (self.attn_o_tnh, P(None, None, "model"))
+
+        def _ragged_paged_attention(*args):
+            outputs = ragged_paged_attention(
+                *args,
+                sm_scale=self.scale,
+                q_scale=q_scale,
+                k_scale=k_scale,
+                v_scale=v_scale,
+            )
+            return outputs
+
+        output_TNH, kv_cache = jax.jit(
+            jax.shard_map(
+                _ragged_paged_attention,
+                mesh=mesh,
+                in_specs=in_specs,
+                out_specs=out_specs,
+                check_vma=False,
+            ))(
+                q_TNH,
+                k_SKH,
+                v_SKH,
+                kv_cache,
+                md.seq_lens,
+                md.block_tables,
+                md.query_start_loc,
+                md.request_distribution,
+            )
+        return kv_cache, output_TNH
+
+    def mla_attention(
+        self,
+        kv_cache: KVCache,
+        q_TNA: jax.Array,
+        q_rope_TNH: jax.Array,
+        k_SA: jax.Array,
+        k_rope_SH: jax.Array,
+        attention_metadata: AttentionMetadata,
+        mesh: Mesh,
+    ) -> Tuple[KVCache, jax.Array]:
+        """Performs scaled dot-product attention and updates the KV cache.
+
+        This function handles the core attention logic, which varies between
+        prefill and generation modes. In prefill, it computes self-attention
+        over the input sequence with a causal mask. In generation, it attends
+        to the full history of keys and values stored in the cache.
+
+        Args:
+            kv_cache: The key-value cache to be updated and used.
+            q_TNA: Query tensor of shape `(query_seq, num_attention_heads, lkv_dim)`.
+            q_rope_TNH: Query rope tensor of shape `(query_seq, num_attention_heads, rope_dim)`.
+            k_SA: Key tensor of shape `(kv_seq, lkv_dim)`.
+            k_rope_SH: Key rope tensor of shape `(kv_seq, rope_dim)`.
+            attention_metadata: Metadata containing sequence lengths.
+            mesh: The JAX device mesh (unused in this specific function but
+                kept for potential future use or API consistency).
+            q_scale: Quantization scale for q.
+            k_scale: Quantization scale for k.
+            v_scale: Quantization scale for v.
+
+        Returns:
+            A tuple containing:
+                - The updated KV cache.
+                - The attention output tensor of shape
+                  `(seq, num_q_heads, head_dim)`.
+        """
+        md = attention_metadata
+        in_specs = (
+            self.query_tnh,  # q
+            self.query_tnh,  # q_rope
+            self.keyvalue_skh,  # k
+            self.keyvalue_skh,  # k_rope
+            P(ShardingAxisName.MLP_TENSOR),  # kv_cache
+            P(ShardingAxisName.ATTN_DATA),  # md.seq_lens: Replicated
+            P(ShardingAxisName.ATTN_DATA),  # page_indices_flat: Replicated
+            P(ShardingAxisName.ATTN_DATA),  # query_start_loc: Replicated
+            P(ShardingAxisName.ATTN_DATA),  # distribution: Replicated
+        )
+
+        out_specs = (self.attn_o_tnh, P(ShardingAxisName.MLP_TENSOR))
+
+        def _mla_ragged_paged_attention(q, q_rope, k, k_rope, kv_cache, *args):
+
+            def _initialize_block_sizes():
+                # Set reasonable starting estimates for block sizes. (TODO(gpolovets): update this to use tuned sizes)
+                # Referring to get_tuned_block_sizes() in kernels/ragged_paged_attention/v3/tuned_block_sizes.py: 'TPU v7'/128/'q_bfloat16_kv_bfloat16/q_head-128_kv_head-1_head-128'/4096
+                max_num_tokens = q.shape[0]
+                max_num_seqs = md.seq_lens.shape[0]
+                num_page_indices = md.block_tables.shape[0]
+                assert num_page_indices % max_num_seqs == 0
+                pages_per_seq = num_page_indices // max_num_seqs
+                # num_kv_pages_per_block = min(pages_per_seq, 16)
+                bkv_p, bq_sz = get_tuned_block_sizes(
+                    q.dtype,
+                    kv_cache.dtype,
+                    self.num_attention_heads,
+                    1,
+                    self.qk_nope_head_dim,
+                    kv_cache.shape[1],  # page size
+                    max_num_tokens,
+                    pages_per_seq,
+                )
+                num_kv_pages_per_block = min(min(pages_per_seq, bkv_p), 4)
+                num_queries_per_block = min(min(max_num_tokens, bq_sz),
+                                            4)  # OOMS at 8
+                return num_kv_pages_per_block, num_queries_per_block
+
+            num_kv_pages_per_block, num_queries_per_block = _initialize_block_sizes(
+            )
+            output, kv_cache = mla_ragged_paged_attention(
+                q,
+                q_rope,
+                k,
+                k_rope,
+                kv_cache,
+                *args,
+                sm_scale=self.scale,
+                num_kv_pages_per_block=num_kv_pages_per_block,
+                num_queries_per_block=num_queries_per_block)
+
+            return kv_cache, output
+
+        kv_cache, output_TNH = jax.jit(
+            jax.shard_map(
+                _mla_ragged_paged_attention,
+                mesh=mesh,
+                in_specs=in_specs,
+                out_specs=out_specs,
+                check_vma=False,
+            ), )(
+                q_TNA,
+                q_rope_TNH,
+                k_SA,
+                k_rope_SH,
+                kv_cache,
+                md.seq_lens,
+                md.block_tables,
+                md.query_start_loc,
+                md.request_distribution,
+            )
+        return kv_cache, output_TNH