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

tpu_inference/layers/vllm/fused_moe.py

@@ -0,0 +1,502 @@
+# 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 functools
+
+import jax
+from jax import numpy as jnp
+from jax.sharding import Mesh
+from jax.sharding import PartitionSpec as P
+
+from tpu_inference.kernels.megablox.gmm import gmm
+from tpu_inference.layers.vllm.linear_common import \
+    slice_sharded_tensor_for_concatenation
+
+
+def activation_fn(activation: str, x1: jax.Array, x2: jax.Array) -> jax.Array:
+    match activation:
+        case "silu":
+            return jax.nn.silu(x1) * x2
+        case "swigluoai":
+            return _swigluoai(x1, x2)
+        case _:
+            raise NotImplementedError(
+                f"FusedMoE does not support {activation} activation")
+
+
+def _swigluoai(x1: jax.Array,
+               x2: jax.Array,
+               alpha=1.702,
+               limit=7.0) -> jax.Array:
+    x1 = jnp.clip(x1, a_max=limit)
+    x2 = jnp.clip(x2, a_min=-limit, a_max=limit)
+
+    gated_activation = x1 * jax.nn.sigmoid(alpha * x1)
+
+    return gated_activation * (x2 + 1)
+
+
+def _round_up_to_multiple_of_128_within_limit(x: int, limit: int) -> int:
+    """
+    Rounds the given integer `x` up to the nearest multiple of 128, without
+    exceeding the specified `limit`.
+
+    If `x` is less than or equal to 128, returns 128.
+    If `x` is less than `limit`, returns the smallest multiple of 128 greater
+    than or equal to `x`.
+    If `x` is greater than or equal to `limit`, searches for the largest
+    multiple of 128 less than or equal to `limit` (down to 512) that divides `x`
+    evenly, and returns it.
+    If no such candidate is found, returns `limit`.
+
+    Args:
+        x (int): The integer to round up.
+        limit (int): The upper bound (must be a multiple of 128).
+
+    Returns:
+        int: The rounded value according to the rules above.
+
+    Raises:
+        AssertionError: If `limit` is less than 128 or not a multiple of 128.
+    """
+    assert limit >= 128 and limit % 128 == 0
+    if x <= 128:
+        return 128
+    if x < limit:
+        return (x + 127) // 128 * 128
+    for candidate in range(limit, 511, -128):
+        if x % candidate == 0:
+            return candidate
+    return limit
+
+
+def _get_tiling_size_for_gmm_kernel(m: int, k: int, n: int,
+                                    g: int) -> tuple[int, int, int]:
+    """
+    Calculate optimal tiling sizes for a GMM kernel in a Mixture of Experts
+    (MoE) setting.
+
+    Args:
+        m (int): The total number of tokens.
+        n (int): The output feature dimension.
+        k (int): The input feature dimension.
+        g (int): The number of experts.
+
+    Returns:
+        tuple[int, int, int]: A tuple (tm, tk, tn)
+    """
+
+    # TODO(Chengji): increase the upper limit tiling size of m when we can set
+    # the vmem size to be used for gmm kernel.
+    # NOTE: In average each expert has m // g tokens, but as it might be
+    # unbalanced, here we doubled the token size when choosing tiling size of m.
+    # 2m//g can be either greater or less than 512. If there are 32 tokens and
+    # topk=2, m=topk * num_tokens=64, in this case, 2*m//g will be less than
+    # 512.
+    tm = _round_up_to_multiple_of_128_within_limit(2 * m // g, 512)
+    tm = min(tm, m)  # there's a requirement that m % tm == 0
+    # k/n correspond to n_input_features/n_output_features in the matmul so they
+    # are normally greater than 2048, unless the num shards is large.
+    tk = _round_up_to_multiple_of_128_within_limit(k, 2048)
+    tn = _round_up_to_multiple_of_128_within_limit(n, 2048)
+    return tm, tk, tn
+
+
+def tensor_sharded_gmm_merged_column_parallel(
+    lhs: jax.Array,
+    rhs: jax.Array,
+    rhs_scale: jax.Array | None,
+    rhs_bias: jax.Array | None,
+    group_sizes: jax.Array,
+    mesh: Mesh,
+) -> list[jax.Array]:
+
+    def _gmm(lhs, rhs, rhs_scale, rhs_bias, group_sizes):
+        m, g, n, k = lhs.shape[0], *rhs.shape
+        tm, tk, tn = _get_tiling_size_for_gmm_kernel(m, k, n, g)
+        return gmm(
+            lhs,
+            rhs,
+            group_sizes,
+            rhs_scale=rhs_scale,
+            rhs_bias=rhs_bias,
+            preferred_element_type=lhs.dtype,
+            tiling=(tm, tk, tn),
+            transpose_rhs=True,
+            group_offset=jnp.array(0),
+        )
+
+    rhs_scale_spec = None if rhs_scale is None else P(None, None, None,
+                                                      "model")
+    rhs_bias_spec = None if rhs_bias is None else P(None, None, "model")
+
+    gmm_result = jax.shard_map(
+        _gmm,
+        mesh=mesh,
+        in_specs=(
+            P("data", None),
+            P(None, "model", None),
+            rhs_scale_spec,
+            rhs_bias_spec,
+            P("data"),
+        ),
+        out_specs=(P("data", "model")),
+        check_vma=False,
+    )(lhs, rhs, rhs_scale, rhs_bias, group_sizes)
+
+    tp_size = mesh.shape["model"]
+    intermediate_size = gmm_result.shape[-1] // 2
+    output_sizes = [intermediate_size, intermediate_size]
+    return slice_sharded_tensor_for_concatenation(gmm_result, output_sizes,
+                                                  tp_size)
+
+
+def tensor_sharded_gmm_row_parallel(
+    lhs: jax.Array,
+    rhs: jax.Array,
+    rhs_scale: jax.Array | None,
+    rhs_bias: jax.Array | None,
+    group_sizes: jax.Array,
+    mesh: Mesh,
+) -> jax.Array:
+
+    def _gmm_all_reduce(lhs, rhs, rhs_scale, rhs_bias, group_sizes):
+        m, g, n, k = lhs.shape[0], *rhs.shape
+        tm, tk, tn = _get_tiling_size_for_gmm_kernel(m, k, n, g)
+        if rhs_bias is not None:
+            shard_id = jax.lax.axis_index("model")
+            rhs_bias = jnp.where(shard_id == 0, rhs_bias, 0)
+        out = gmm(
+            lhs,
+            rhs,
+            group_sizes,
+            rhs_scale=rhs_scale,
+            rhs_bias=rhs_bias,
+            preferred_element_type=lhs.dtype,
+            tiling=(tm, tk, tn),
+            transpose_rhs=True,
+            group_offset=jnp.array(0),
+        )
+        return jax.lax.psum(out, axis_name="model")
+
+    num_blocks = 1 if rhs_scale is None else rhs_scale.shape[1]
+    rhs_scale_spec = None if num_blocks == 1 else P(None, "model", None, None)
+    rhs_bias_spec = None if rhs_bias is None else P(None, None, None)
+    gmm_result = jax.shard_map(
+        _gmm_all_reduce,
+        mesh=mesh,
+        in_specs=(
+            P("data", "model"),
+            P(None, None, "model"),
+            rhs_scale_spec,
+            rhs_bias_spec,
+            P("data"),
+        ),
+        out_specs=(P("data")),
+        check_vma=False,
+    )(lhs, rhs, rhs_scale, rhs_bias, group_sizes)
+
+    return gmm_result.astype(lhs.dtype)
+
+
+def expert_sharded_gmm(
+    lhs: jax.Array,
+    rhs: jax.Array,
+    rhs_scale: jax.Array | None,
+    rhs_bias: jax.Array | None,
+    group_sizes: jax.Array,
+    is_last_expert: bool,
+    mesh: Mesh,
+) -> jax.Array:
+    ep_size = mesh.shape["model"]
+
+    num_experts = rhs.shape[0]
+    num_experts_per_shard = num_experts // ep_size
+    group_offset = jnp.arange(0, num_experts, num_experts_per_shard)
+
+    def _gmm(lhs, rhs, rhs_scale, rhs_bias, group_sizes, group_offset):
+        m, g, n, k = lhs.shape[0], *rhs.shape
+        tm, tk, tn = _get_tiling_size_for_gmm_kernel(m, k, n, g)
+
+        gmm_res = gmm(
+            lhs=lhs,
+            rhs=rhs,
+            rhs_scale=rhs_scale,
+            rhs_bias=rhs_bias,
+            group_sizes=group_sizes,
+            preferred_element_type=lhs.dtype,
+            tiling=(tm, tk, tn),
+            transpose_rhs=True,
+            group_offset=group_offset[0],
+        )
+        return gmm_res
+
+    # The result from gmm on each shard has the same shape, but only the rows
+    # for this shard has non-zero values. Taking below as an working example:
+    #       A, A, A, A     0, 0, 0, 0     0, 0, 0, 0     0, 0, 0, 0
+    #       A, A, A, A     0, 0, 0, 0     0, 0, 0, 0     0, 0, 0, 0
+    #       A, A, A, A     0, 0, 0, 0     0, 0, 0, 0     0, 0, 0, 0
+    #       0, 0, 0, 0     B, B, B, B     0, 0, 0, 0     0, 0, 0, 0
+    #       0, 0, 0, 0     B, B, B, B     0, 0, 0, 0     0, 0, 0, 0
+    #       0, 0, 0, 0     0, 0, 0, 0     C, C, C, C     0, 0, 0, 0
+    #       0, 0, 0, 0     0, 0, 0, 0     C, C, C, C     0, 0, 0, 0
+    #       0, 0, 0, 0     0, 0, 0, 0     C, C, C, C     0, 0, 0, 0
+    #       0, 0, 0, 0     0, 0, 0, 0     C, C, C, C     0, 0, 0, 0
+    #       0, 0, 0, 0     0, 0, 0, 0     C, C, C, C     0, 0, 0, 0
+    #       0, 0, 0, 0     0, 0, 0, 0     0, 0, 0, 0     D, D, D, D
+    #       0, 0, 0, 0     0, 0, 0, 0     0, 0, 0, 0     D, D, D, D
+    #       0, 0, 0, 0     0, 0, 0, 0     0, 0, 0, 0     D, D, D, D
+    #       0, 0, 0, 0     0, 0, 0, 0     0, 0, 0, 0     D, D, D, D
+    #        shard-0        shard-1        shard-2        shard-3
+    # Each shards has 3 (row A), 2 (row B), 5 (row C) and 4 (row D).
+    lhs_spec = P("model") if is_last_expert else P()
+    rhs_scale_spec = None if rhs_scale is None else P("model")
+    rhs_bias_spec = None if rhs_bias is None else P("model")
+    gmm_res = jax.shard_map(
+        _gmm,
+        mesh=mesh,
+        in_specs=(
+            lhs_spec,
+            P("model", None, None),
+            rhs_scale_spec,
+            rhs_bias_spec,
+            P(),
+            P("model"),
+        ),
+        out_specs=(P("model", None)),
+        check_vma=False,
+    )(lhs, rhs, rhs_scale, rhs_bias, group_sizes, group_offset)
+
+    if not is_last_expert:
+        return gmm_res
+
+    # For i-th shard, it is responsible groups (AKA experts) from
+    # i*num_experts_per_shard to (i+1)*num_experts_per_shard We sum them up to
+    # get total rows in that shard, and that is the size for shard to send to
+    # its peers. This is also the number of non-zero rows from the gmm results.
+    # In the working example, send_sizes would be [3, 2, 5, 4].
+
+    # group_sizes has shape of [num_tokens_per_shard * num_experts_per_shard].
+    # So reshaping to [num_tokens_per_shard, num_experts_per_shard] and applying
+    # sum(axis=1) will get desired send_sizes shaped [num_tokens_per_shard].
+    send_sizes = group_sizes.reshape(-1, num_experts_per_shard).sum(axis=1)
+    # In the working example, input_offsets would be [0, 3, 5, 10]
+    input_offsets = jnp.concatenate((jnp.array([0]), send_sizes.cumsum()[:-1]))
+    output_offsets = input_offsets
+    recv_sizes = send_sizes
+
+    def _ragged_all_to_all(operand, input_offsets, send_sizes, output_offsets,
+                           recv_sizes):
+        output = jnp.zeros_like(operand)
+
+        # input_offsets, send_sizes and output_offsets are sharded and there is
+        # only 1 elemnt in each shard, we are taking the 0-th element from them
+        # just so that jnp.repeat generates the arrays with correct shape.
+        input_offsets_of_shard = jnp.repeat(input_offsets[0], ep_size)
+        send_sizes_of_shard = jnp.repeat(send_sizes[0], ep_size)
+        output_offsets_of_shard = jnp.repeat(output_offsets[0], ep_size)
+
+        # recv_sizes is replicated across shards, because all the shards receive
+        # the same data and write to the output in the same way (same
+        # output_offsets and same recv_sizes) and thus generates replicated
+        # output.
+        recv_sizes_of_shard = recv_sizes
+
+        # In the working example, for each shard, the values of the offsets and
+        # sizes would be:
+        #                                shard-0         shard-1         shard-2         shard-3
+        # input_offsets_of_shard       [0, 0, 0, 0]    [3, 3, 3, 3]    [5, 5, 5, 5]    [10,10,10,10]
+        # send_sizes_of_shard          [3, 3, 3, 3]    [2, 2, 2, 2]    [5, 5, 5, 5]    [4, 4, 4, 4 ]
+        # output_offsets_of_shard      [0, 0, 0, 0]    [0, 0, 0, 0]    [0, 0, 0, 0]    [10,10,10,10]
+        # recv_sizes_of_shard          [3, 2, 5, 4]    [3, 2, 5, 4]    [3, 2, 5, 4]    [3, 2, 5, 4]
+        return jax.lax.ragged_all_to_all(
+            operand,
+            output,
+            input_offsets_of_shard,
+            send_sizes_of_shard,
+            output_offsets_of_shard,
+            recv_sizes_of_shard,
+            axis_name="model",
+        )
+
+    # Use ragged_all_to_all to send the result from gmm for each expert to all
+    # the shards.  In the working example, the result would be:
+    #       A, A, A, A     A, A, A, A     A, A, A, A     A, A, A, A
+    #       A, A, A, A     A, A, A, A     A, A, A, A     A, A, A, A
+    #       A, A, A, A     A, A, A, A     A, A, A, A     A, A, A, A
+    #       B, B, B, B     B, B, B, B     B, B, B, B     B, B, B, B
+    #       B, B, B, B     B, B, B, B     B, B, B, B     B, B, B, B
+    #       C, C, C, C     C, C, C, C     C, C, C, C     C, C, C, C
+    #       C, C, C, C     C, C, C, C     C, C, C, C     C, C, C, C
+    #       C, C, C, C     C, C, C, C     C, C, C, C     C, C, C, C
+    #       C, C, C, C     C, C, C, C     C, C, C, C     C, C, C, C
+    #       C, C, C, C     C, C, C, C     C, C, C, C     C, C, C, C
+    #       D, D, D, D     D, D, D, D     D, D, D, D     D, D, D, D
+    #       D, D, D, D     D, D, D, D     D, D, D, D     D, D, D, D
+    #       D, D, D, D     D, D, D, D     D, D, D, D     D, D, D, D
+    #       D, D, D, D     D, D, D, D     D, D, D, D     D, D, D, D
+    #        shard-0        shard-1        shard-2        shard-3
+    return jax.shard_map(
+        _ragged_all_to_all,
+        mesh=mesh,
+        in_specs=(P("model", None), P("model"), P("model"), P("model"), P()),
+        out_specs=(P()),
+        check_vma=False,
+    )(gmm_res, input_offsets, send_sizes, output_offsets, recv_sizes)
+
+
+@functools.partial(
+    jax.jit,
+    static_argnames=(
+        "topk",
+        "renormalize",
+        "mesh",
+        "use_ep",
+        "activation",
+    ),
+)
+def fused_moe_func(
+    hidden_states: jax.Array,
+    w1: jax.Array,
+    w2: jax.Array,
+    w1_scale: jax.Array | None,
+    w2_scale: jax.Array | None,
+    w1_bias: jax.Array | None,
+    w2_bias: jax.Array | None,
+    gating_output: jax.Array,
+    topk: int,
+    renormalize: bool,
+    mesh: Mesh,
+    use_ep: bool,
+    activation: str,
+) -> jax.Array:
+    """Route tokens in hidden_states into each experts based on routing.
+
+    Args:
+        hidden_states: [num_tokens, hidden_size]
+        w1: first moe weights [num_experts, intermediate_size * 2, hidden_size]
+        w2: second moe weights [num_experts, hidden_size, intermediate_size]
+        w1_scale: w1 scale [num_experts, num_blocks, 1, intermediate_size * 2]
+        w2_scale: w2 scale [num_experts, num_blocks, 1, hidden_size]
+        w1_bias: optional bias of w1 [num_experts, 1, intermediate_size * 2]
+        w2_bias: optional bias of w2 [num_experts, 1, hidden_size]
+        gating_output: routing information of tokens [num_tokens, num_experts]
+        topk: number of experts to choose per token.
+        renormalize: normalize gating_output.
+        mesh: mesh to perform moe.
+        use_ep: use expert parallelism.
+        activation: activation function to perform on the output of w1.
+
+    Returns:
+        Output of moe operation [num_tokens, hidden_size]
+    """
+    num_tokens, hidden_size = hidden_states.shape
+    global_num_experts, _, padded_hidden_size = w1.shape
+    dtype = hidden_states.dtype
+
+    assert (num_tokens * topk) % 16 == 0, (
+        "The kernel requires num_tokens * topk to be a multiple of "
+        f"16 but got {num_tokens}*{topk}={num_tokens*topk}")
+
+    assert gating_output.shape == (num_tokens, global_num_experts)
+
+    topk_weights = jax.nn.softmax(gating_output.astype(jnp.float32), axis=-1)
+    topk_weights, topk_indices = jax.lax.top_k(topk_weights, k=topk)
+    if renormalize:
+        topk_weights = topk_weights / topk_weights.sum(axis=-1, keepdims=True)
+    topk_weights = topk_weights.astype(dtype)
+
+    def _process_tokens_locally(hidden_states_local, topk_indices_local):
+        num_tokens_local = hidden_states_local.shape[0]
+        topk_indices_flat = topk_indices_local.flatten()
+        topk_argsort_indices = jnp.argsort(topk_indices_flat)
+        topk_argsort_revert_indices = jnp.argsort(topk_argsort_indices)
+        token_indices = jnp.arange(num_tokens_local,
+                                   dtype=jnp.int32).repeat(topk)
+        token_indices_sorted = token_indices[topk_argsort_indices]
+        group_sizes_local = jnp.bincount(topk_indices_flat,
+                                         length=global_num_experts)
+
+        x = hidden_states_local[token_indices_sorted]
+        return x, group_sizes_local, topk_argsort_revert_indices
+
+    x, group_sizes, topk_argsort_revert_indices = jax.shard_map(
+        _process_tokens_locally,
+        mesh=mesh,
+        in_specs=(P("data", None), P("data", None)),
+        out_specs=(P("data", None), P("data"), P("data")),
+    )(hidden_states, topk_indices)
+
+    x = jnp.pad(x, ((0, 0), (0, padded_hidden_size - hidden_size)))
+
+    if use_ep:
+        x = expert_sharded_gmm(
+            x,
+            w1,
+            w1_scale,
+            w1_bias,
+            group_sizes,
+            is_last_expert=False,
+            mesh=mesh,
+        )
+        x1, x2 = jnp.split(x, 2, -1)
+
+        x = activation_fn(activation, x1, x2)
+
+        x = expert_sharded_gmm(
+            x,
+            w2,
+            w2_scale,
+            w2_bias,
+            group_sizes,
+            is_last_expert=True,
+            mesh=mesh,
+        )
+    else:
+        x1, x2 = tensor_sharded_gmm_merged_column_parallel(
+            x,
+            w1,
+            w1_scale,
+            w1_bias,
+            group_sizes,
+            mesh=mesh,
+        )
+
+        x = activation_fn(activation, x1, x2)
+
+        x = tensor_sharded_gmm_row_parallel(
+            x,
+            w2,
+            w2_scale,
+            w2_bias,
+            group_sizes,
+            mesh=mesh,
+        )
+
+    def _finalize_output(x_local, topk_argsort_revert_indices_local,
+                         topk_weights_local):
+        x_local = x_local[topk_argsort_revert_indices_local].reshape(
+            -1, topk, padded_hidden_size)
+        x_local = x_local * jnp.expand_dims(topk_weights_local, axis=-1)
+        x_local = x_local.sum(axis=-2)
+        return x_local
+
+    x = jax.shard_map(
+        _finalize_output,
+        mesh=mesh,
+        in_specs=(P("data", None), P("data"), P("data", None)),
+        out_specs=(P("data", None)),
+    )(x, topk_argsort_revert_indices, topk_weights)
+
+    return x[:num_tokens, :hidden_size]