tpu-inference 0.11.1.dev202511270815__py3-none-any.whl → 0.13.0rc2.post7__py3-none-any.whl

tpu_inference/layers/vllm/fused_moe.py

@@ -1,18 +1,32 @@
+# 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.experimental.pallas.ops.tpu.megablox.gmm import gmm
-from jax.experimental.shard_map import shard_map
-from jax.sharding import Mesh, NamedSharding, PartitionSpec
+from jax.sharding import Mesh, NamedSharding
+from jax.sharding import PartitionSpec as P
 
+from tpu_inference.kernels.megablox.gmm import gmm
+from tpu_inference.layers.common.sharding import ShardingAxisName
 from tpu_inference.layers.vllm.linear_common import \
     slice_sharded_tensor_for_concatenation
-
-P = PartitionSpec
+from tpu_inference.utils import get_mesh_shape_product
 
 
-def activation_fn(activation: str, x1, x2):
+def activation_fn(activation: str, x1: jax.Array, x2: jax.Array) -> jax.Array:
     match activation:
         case "silu":
             return jax.nn.silu(x1) * x2
@@ -23,7 +37,10 @@ def activation_fn(activation: str, x1, x2):
                 f"FusedMoE does not support {activation} activation")
 
 
-def _swigluoai(x1, x2, alpha=1.702, limit=7.0):
+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)
 
@@ -101,142 +118,124 @@ def _get_tiling_size_for_gmm_kernel(m: int, k: int, n: int,
 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,
-    transpose_rhs: bool,
     mesh: Mesh,
-    intermediate_size: int,
-) -> jax.Array:
-    # adapted from https://github.com/pytorch/xla/blob/1d409399474197c484894be90b75d9855393dda5/torch_xla/experimental/custom_kernel.py#L1401
-    m, k, g = lhs.shape[0], lhs.shape[1], rhs.shape[0]
-    n = rhs.shape[1] if transpose_rhs else rhs.shape[2]
-    tm, tk, tn = _get_tiling_size_for_gmm_kernel(m // mesh.shape["data"], k, n,
-                                                 g)
-
-    _gmm = functools.partial(
-        gmm,
-        preferred_element_type=lhs.dtype,
-        tiling=(tm, tk, tn),
-        transpose_rhs=transpose_rhs,
-        group_offset=jnp.array(0),
-    )
-
-    gmm_result = shard_map(
+) -> 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, ShardingAxisName.MLP_TENSOR)
+    rhs_bias_spec = None if rhs_bias is None else P(
+        None, None, ShardingAxisName.MLP_TENSOR)
+
+    gmm_result = jax.shard_map(
         _gmm,
         mesh=mesh,
-        in_specs=(P("data", None), P(None, "model", None), P("data")),
-        out_specs=(P("data", "model")),
-        check_rep=False,
-    )(lhs, rhs, group_sizes)
-
-    if rhs_bias is not None:
-
-        def _add_bias(gmm_result_local, rhs_bias_local, group_sizes_global):
-            rhs_bis = jnp.repeat(rhs_bias_local,
-                                 group_sizes_global,
-                                 0,
-                                 total_repeat_length=m // mesh.shape["data"])
-            return (gmm_result_local + rhs_bis).astype(gmm_result_local.dtype)
-
-        gmm_result = shard_map(
-            _add_bias,
-            mesh=mesh,
-            in_specs=(P("data", "model"), P(None, "model"), P("data")),
-            out_specs=(P("data", "model")),
-        )(gmm_result, rhs_bias, group_sizes)
-
-    n_shards = mesh.shape["model"]
+        in_specs=(P(ShardingAxisName.MLP_DATA,
+                    None), P(None, ShardingAxisName.MLP_TENSOR,
+                             None), rhs_scale_spec, rhs_bias_spec,
+                  P(ShardingAxisName.MLP_DATA)),
+        out_specs=(P(ShardingAxisName.MLP_DATA, ShardingAxisName.MLP_TENSOR)),
+        check_vma=False,
+    )(lhs, rhs, rhs_scale, rhs_bias, group_sizes)
+
+    tp_size = get_mesh_shape_product(mesh, ShardingAxisName.MLP_TENSOR)
+    intermediate_size = gmm_result.shape[-1] // 2
     output_sizes = [intermediate_size, intermediate_size]
-
     return slice_sharded_tensor_for_concatenation(gmm_result, output_sizes,
-                                                  n_shards)
+                                                  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,
-    transpose_rhs: bool,
     mesh: Mesh,
 ) -> jax.Array:
-    # adapted from https://github.com/pytorch/xla/blob/1d409399474197c484894be90b75d9855393dda5/torch_xla/experimental/custom_kernel.py#L1401
-    m, k, g = lhs.shape[0], lhs.shape[1], rhs.shape[0]
-    n = rhs.shape[1] if transpose_rhs else rhs.shape[2]
-    tm, tk, tn = _get_tiling_size_for_gmm_kernel(m // mesh.shape["data"], k, n,
-                                                 g)
-
-    _gmm = functools.partial(
-        gmm,
-        preferred_element_type=lhs.dtype,
-        tiling=(tm, tk, tn),
-        transpose_rhs=transpose_rhs,
-        group_offset=jnp.array(0),
-    )
-
-    def _gmm_all_reduce(lhs, rhs, group_sizes):
-        r = _gmm(lhs, rhs, group_sizes)
-        return jax.lax.psum(r, axis_name="model")
-
-    gmm_result = shard_map(
+
+    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(ShardingAxisName.MLP_TENSOR).sum()
+            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=ShardingAxisName.MLP_TENSOR)
+
+    num_blocks = 1 if rhs_scale is None else rhs_scale.shape[1]
+    rhs_scale_spec = None if num_blocks == 1 else P(
+        None, ShardingAxisName.MLP_TENSOR, 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"), P("data")),
-        out_specs=(P("data")),
-        check_rep=False,
-    )(lhs, rhs, group_sizes)
-    if rhs_bias is not None:
-
-        def _add_bias(gmm_result_local, rhs_bias_local, group_sizes_global):
-            rhs_bis = jnp.repeat(rhs_bias_local,
-                                 group_sizes_global,
-                                 0,
-                                 total_repeat_length=m // mesh.shape["data"])
-            return (gmm_result_local + rhs_bis).astype(gmm_result_local.dtype)
-
-        gmm_result = shard_map(
-            _add_bias,
-            mesh=mesh,
-            in_specs=(P("data"), P(), P("data")),
-            out_specs=(P("data")),
-        )(gmm_result, rhs_bias, group_sizes)
+        in_specs=(P(ShardingAxisName.MLP_DATA, ShardingAxisName.MLP_TENSOR),
+                  P(None, None, ShardingAxisName.MLP_TENSOR), rhs_scale_spec,
+                  rhs_bias_spec, P(ShardingAxisName.MLP_DATA)),
+        out_specs=(P(ShardingAxisName.MLP_DATA)),
+        check_vma=False,
+    )(lhs, rhs, rhs_scale, rhs_bias, group_sizes)
 
-    return gmm_result
+    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,
-    transpose_rhs: bool,
+    is_last_expert: bool,
     mesh: Mesh,
-    num_experts: int,
-    ep_size: int,
 ) -> jax.Array:
-    # adapted from https://github.com/pytorch/xla/blob/1d409399474197c484894be90b75d9855393dda5/torch_xla/experimental/custom_kernel.py#L1401
-    m, k, g = lhs.shape[0], lhs.shape[1], rhs.shape[0]
-    n = rhs.shape[1] if transpose_rhs else rhs.shape[2]
-    tm, tk, tn = _get_tiling_size_for_gmm_kernel(m, k, n, g)
-
+    ep_size = get_mesh_shape_product(mesh, ShardingAxisName.MLP_TENSOR)
+    ep_p_spec = P(ShardingAxisName.EXPERT)
+    num_experts = rhs.shape[0]
     num_experts_per_shard = num_experts // ep_size
     group_offset = jnp.arange(0, num_experts, num_experts_per_shard)
-    group_offset = jax.lax.with_sharding_constraint(
-        group_offset, NamedSharding(mesh, P("model")))
-
-    def _gmm(lhs, rhs, group_sizes, group_offset):
-        # Group offset for this shard. `group_offset` is sharded, and in this
-        # sharded function, it has only 1 element and `group_offset.shape` is
-        # (1,) but gmm kernel requires the group_offset to be a ()-shaped array,
-        # so we group_offset[0].
-        group_offset_of_shard = group_offset[0]
+
+    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=transpose_rhs,
-            group_offset=group_offset_of_shard,
+            transpose_rhs=True,
+            group_offset=group_offset[0],
         )
         return gmm_res
 
@@ -258,35 +257,43 @@ def expert_sharded_gmm(
     #       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).
-    gmm_res = shard_map(
+    lhs_spec = ep_p_spec if is_last_expert else P()
+    rhs_spec = ep_p_spec
+    rhs_scale_spec = None if rhs_scale is None else ep_p_spec
+    rhs_bias_spec = None if rhs_bias is None else ep_p_spec
+    gmm_res = jax.shard_map(
         _gmm,
         mesh=mesh,
-        in_specs=(P(), P("model", None, None), P(), P("model")),
-        out_specs=(P("model", None)),
-        check_rep=False,
-    )(lhs, rhs, group_sizes, group_offset)
+        in_specs=(
+            lhs_spec,
+            rhs_spec,
+            rhs_scale_spec,
+            rhs_bias_spec,
+            P(),
+            ep_p_spec,
+        ),
+        out_specs=ep_p_spec,
+        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]
-    send_sizes = jnp.array([
-        group_sizes[i * num_experts_per_shard:(i + 1) *
-                    num_experts_per_shard].sum() for i in range(ep_size)
-    ])
+    # 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
 
-    input_offsets = jax.lax.with_sharding_constraint(
-        input_offsets, NamedSharding(mesh, P("model")))
-    send_sizes = jax.lax.with_sharding_constraint(
-        send_sizes, NamedSharding(mesh, P("model")))
-    output_offsets = jax.lax.with_sharding_constraint(
-        output_offsets, NamedSharding(mesh, P("model")))
-
     def _ragged_all_to_all(operand, input_offsets, send_sizes, output_offsets,
                            recv_sizes):
         output = jnp.zeros_like(operand)
@@ -317,7 +324,7 @@ def expert_sharded_gmm(
                                          send_sizes_of_shard,
                                          output_offsets_of_shard,
                                          recv_sizes_of_shard,
-                                         axis_name="model")
+                                         axis_name=ShardingAxisName.EXPERT)
 
     # 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:
@@ -336,56 +343,74 @@ def expert_sharded_gmm(
     #       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 shard_map(
+    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_rep=False,
+        in_specs=(ep_p_spec, ep_p_spec, ep_p_spec, ep_p_spec, P()),
+        out_specs=(P(ShardingAxisName.MLP_DATA)),
+        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,
-    global_num_experts: int,
     renormalize: bool,
-    reduce_results: bool,
     mesh: Mesh,
     use_ep: bool,
     activation: str,
-):
-    """
+) -> jax.Array:
+    """Route tokens in hidden_states into each experts based on routing.
+
     Args:
-        hidden_states: [*, hidden_size]
-        w1: [num_experts, intermediate_size * 2, hidden_size]
-        w2: [num_experts, hidden_size, intermediate_size]
-        gating_output: [*, num_experts]
+        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]
     """
-    # adapted from https://github.com/vllm-project/vllm/blob/29fa5cac1cd731026f59084d93a822921507573c/vllm/model_executor/layers/fused_moe/moe_pallas.py#L26
-    if use_ep and (w1_bias is not None or w2_bias is not None):
-        raise NotImplementedError(
-            "Bias is not supported when using expert parallelism.")
-    orig_shape = hidden_states.shape
-    hidden_size = hidden_states.shape[-1]
-    num_tokens = hidden_states.size // hidden_size
-    assert global_num_experts == w1.shape[0]
-    ep_size = mesh.shape["model"]  # only used if use_ep is True.
-    intermediate_size = w2.shape[-1]
+    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}")
 
-    hidden_states = hidden_states.reshape(num_tokens, hidden_size)
-    gating_output = gating_output.reshape(num_tokens, global_num_experts)
+    assert gating_output.shape == (num_tokens, global_num_experts)
 
     topk_weights = jax.nn.softmax(gating_output.astype(jnp.float32), axis=-1)
+    # All-gather topk weights for attention dp
+    topk_weights = jax.lax.with_sharding_constraint(
+        topk_weights, NamedSharding(mesh, P(ShardingAxisName.MLP_DATA, None)))
     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)
@@ -405,152 +430,77 @@ def fused_moe_func(
         x = hidden_states_local[token_indices_sorted]
         return x, group_sizes_local, topk_argsort_revert_indices
 
-    x, group_sizes, topk_argsort_revert_indices = shard_map(
+    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")),
-        check_rep=False,
+        in_specs=(P(ShardingAxisName.MLP_DATA,
+                    None), P(ShardingAxisName.MLP_DATA, None)),
+        out_specs=(P(ShardingAxisName.MLP_DATA, None),
+                   P(ShardingAxisName.MLP_DATA), P(ShardingAxisName.MLP_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,
-            group_sizes,
-            transpose_rhs=True,
-            mesh=mesh,
-            num_experts=global_num_experts,
-            ep_size=ep_size,
-        )
-        x1, x2 = x[..., :intermediate_size], x[..., intermediate_size:]
-    else:
-        x1, x2 = tensor_sharded_gmm_merged_column_parallel(
-            x,
-            w1,
+            w1_scale,
             w1_bias,
             group_sizes,
-            transpose_rhs=True,
+            is_last_expert=False,
             mesh=mesh,
-            intermediate_size=intermediate_size,
         )
+        x1, x2 = jnp.split(x, 2, -1)
 
-    x = activation_fn(activation, x1, x2)
+        x = activation_fn(activation, x1, x2)
 
-    if use_ep:
         x = expert_sharded_gmm(
             x,
             w2,
+            w2_scale,
+            w2_bias,
             group_sizes,
-            transpose_rhs=True,
+            is_last_expert=True,
             mesh=mesh,
-            num_experts=global_num_experts,
-            ep_size=ep_size,
         )
     else:
-        x = jax.lax.with_sharding_constraint(
-            x, NamedSharding(mesh, P("data", "model")))
+        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,
-            transpose_rhs=True,
             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, hidden_size)
+            -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 = shard_map(
+    x = jax.shard_map(
         _finalize_output,
         mesh=mesh,
-        in_specs=(P("data", None), P("data"), P("data", None)),
-        out_specs=(P("data", None)),
-        check_rep=False,
+        in_specs=(P(ShardingAxisName.MLP_DATA,
+                    None), P(ShardingAxisName.MLP_DATA),
+                  P(ShardingAxisName.MLP_DATA, None)),
+        out_specs=(P(ShardingAxisName.ATTN_DATA, None)),
+        check_vma=False,
     )(x, topk_argsort_revert_indices, topk_weights)
-    x = x.reshape(orig_shape)
 
-    if reduce_results:
-        x = jax.lax.with_sharding_constraint(x, NamedSharding(mesh, P("data")))
-    return x
-
-
-@functools.partial(
-    jax.jit,
-    static_argnames=(
-        "topk",
-        "global_num_experts",
-        "renormalize",
-        "reduce_results",
-        "mesh",
-        "use_ep",
-        "activation",
-    ),
-)
-def fused_moe_func_padded(
-    hidden_states: jax.Array,
-    w1: jax.Array,
-    w2: jax.Array,
-    w1_bias: jax.Array | None,
-    w2_bias: jax.Array | None,
-    gating_output: jax.Array,
-    topk: int,
-    global_num_experts: int,
-    renormalize: bool,
-    reduce_results: bool,
-    mesh: Mesh,
-    use_ep: bool,
-    activation: str,
-):
-    # TODO(fanhongmin@google.com): Once the jax runner pads the input, we no longer need this.
-    hidden_size = hidden_states.shape[-1]
-    num_tokens = hidden_states.size // hidden_size
-    if num_tokens * topk < 16:
-        assert 16 % (num_tokens *
-                     topk) == 0, f"Cannot pad to 16: {num_tokens=}, {topk=}"
-        n_repeats = 16 // (num_tokens * topk)
-
-        reps = (n_repeats, ) + (1, ) * (hidden_states.ndim - 1)
-        expanded_hidden_states = jnp.tile(hidden_states, reps)
-
-        reps = (n_repeats, ) + (1, ) * (gating_output.ndim - 1)
-        expanded_gating_output = jnp.tile(gating_output, reps)
-
-        expanded_x = fused_moe_func(
-            expanded_hidden_states,
-            w1,
-            w2,
-            w1_bias,
-            w2_bias,
-            expanded_gating_output,
-            topk,
-            global_num_experts,
-            renormalize,
-            reduce_results,
-            mesh,
-            use_ep,
-            activation,
-        )
-        x = expanded_x[:hidden_states.shape[0]]
-        return x
-    else:
-        return fused_moe_func(
-            hidden_states,
-            w1,
-            w2,
-            w1_bias,
-            w2_bias,
-            gating_output,
-            topk,
-            global_num_experts,
-            renormalize,
-            reduce_results,
-            mesh,
-            use_ep,
-            activation,
-        )
+    return x[:num_tokens, :hidden_size]