tpu-inference 0.0.1rc1__py3-none-any.whl

tpu_inference/layers/vllm/fused_moe.py

@@ -0,0 +1,469 @@
+import functools
+
+import jax
+from jax import numpy as jnp
+from jax import shard_map
+from jax.experimental.pallas.ops.tpu.megablox.gmm import gmm
+from jax.sharding import Mesh
+from jax.sharding import PartitionSpec as P
+
+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_bias: jax.Array | None,
+    group_sizes: jax.Array,
+    mesh: Mesh,
+) -> tuple[jax.Array, jax.Array]:
+
+    def _gmm(lhs, rhs, 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,
+            preferred_element_type=lhs.dtype,
+            tiling=(tm, tk, tn),
+            transpose_rhs=True,
+            group_offset=jnp.array(0),
+        )
+
+    gmm_result = shard_map(
+        _gmm,
+        mesh=mesh,
+        in_specs=(P("data", None), P(None, "model", None), P("data")),
+        out_specs=(P("data", "model")),
+        check_vma=False,
+    )(lhs, rhs, group_sizes)
+
+    if rhs_bias is not None:
+
+        def _add_bias(gmm_result_local, rhs_bias_local, group_sizes_global):
+            rhs_bias = jnp.repeat(
+                rhs_bias_local,
+                group_sizes_global,
+                0,
+                total_repeat_length=gmm_result_local.shape[0])
+            return gmm_result_local + rhs_bias
+
+        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)
+    gmm_result = gmm_result.astype(lhs.dtype)
+
+    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_bias: jax.Array | None,
+    group_sizes: jax.Array,
+    mesh: Mesh,
+) -> jax.Array:
+
+    def _gmm_all_reduce(lhs, rhs, group_sizes):
+        m, g, n, k = lhs.shape[0], *rhs.shape
+        tm, tk, tn = _get_tiling_size_for_gmm_kernel(m, k, n, g)
+        out = gmm(
+            lhs,
+            rhs,
+            group_sizes,
+            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")
+
+    gmm_result = shard_map(
+        _gmm_all_reduce,
+        mesh=mesh,
+        in_specs=(P("data", "model"), P(None, None, "model"), P("data")),
+        out_specs=(P("data")),
+        check_vma=False,
+    )(lhs, rhs, group_sizes)
+
+    if rhs_bias is not None:
+
+        def _add_bias(gmm_result_local, rhs_bias_local, group_sizes_global):
+            rhs_bias = jnp.repeat(
+                rhs_bias_local,
+                group_sizes_global,
+                0,
+                total_repeat_length=gmm_result_local.shape[0])
+            return gmm_result_local + rhs_bias
+
+        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)
+
+    return gmm_result.astype(lhs.dtype)
+
+
+def expert_sharded_gmm(
+    lhs: jax.Array,
+    rhs: jax.Array,
+    group_sizes: jax.Array,
+    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, 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,
+            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).
+    gmm_res = shard_map(
+        _gmm,
+        mesh=mesh,
+        in_specs=(P(), P("model", None, None), P(), P("model")),
+        out_specs=(P("model", None)),
+        check_vma=False,
+    )(lhs, rhs, group_sizes, group_offset)
+
+    # 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 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_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
+    information in gating_out and performs moe with w1 and w2 weights.
+
+    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_bias: optional bias of w1 [num_experts, intermediate_size * 2]
+        w2_bias: optional bias of w2 [num_experts, 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]
+    """
+    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.")
+
+    num_tokens = hidden_states.shape[0]
+    global_num_experts, hidden_size, intermediate_size = w2.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 hidden_states.shape == (num_tokens, hidden_size)
+    assert gating_output.shape == (num_tokens, global_num_experts)
+    assert w1.shape == (global_num_experts, intermediate_size * 2, hidden_size)
+
+    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 = 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)
+
+    if use_ep:
+        x = expert_sharded_gmm(
+            x,
+            w1,
+            group_sizes,
+            mesh=mesh,
+        )
+        x1, x2 = jnp.split(x, 2, -1)
+
+        x = activation_fn(activation, x1, x2)
+
+        x = expert_sharded_gmm(
+            x,
+            w2,
+            group_sizes,
+            mesh=mesh,
+        )
+    else:
+        x1, x2 = tensor_sharded_gmm_merged_column_parallel(
+            x,
+            w1,
+            w1_bias,
+            group_sizes,
+            mesh=mesh,
+        )
+
+        x = activation_fn(activation, x1, x2)
+
+        x = tensor_sharded_gmm_row_parallel(
+            x,
+            w2,
+            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, 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(
+        _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]

tpu_inference/layers/vllm/linear_common.py

@@ -0,0 +1,186 @@
+from typing import Optional, Union
+
+import jax
+import jax.numpy as jnp
+import torch
+from jax.experimental.shard_map import shard_map
+from jax.sharding import Mesh, NamedSharding
+from jax.sharding import PartitionSpec as P
+from torchax.interop import torch_view
+from torchax.ops.mappings import t2j
+
+from tpu_inference.kernels.quantized_matmul.kernel import \
+    quantized_matmul_kernel
+
+
+def sharded_quantized_matmul(x: jax.Array, w_q: jax.Array, w_s: jax.Array,
+                             mesh: Mesh, weight_sharding: P):
+    out_axis, in_axis = weight_sharding
+    x_sharding = P(None, in_axis)
+    scale_sharding = P(out_axis, )
+    out_sharding = P(None, out_axis)
+
+    x = jax.lax.with_sharding_constraint(x, NamedSharding(mesh, x_sharding))
+
+    def wrapper(x, w_q, w_s):
+        output = quantized_matmul_kernel(x, w_q, w_s, x_q_dtype=w_q.dtype)
+        if in_axis:
+            output = jax.lax.psum(output, axis_name=in_axis)
+        return output
+
+    return shard_map(wrapper,
+                     mesh=mesh,
+                     in_specs=(x_sharding, weight_sharding, scale_sharding),
+                     out_specs=(out_sharding),
+                     check_rep=False)(x, w_q, w_s)
+
+
+def reorder_concatenated_tensor_for_sharding(concatenated_tensor: jax.Array,
+                                             split_sizes: list[int],
+                                             n_shards: int, dim: int):
+    """
+    Reorder a replicated concatenated tensor such that when sharded on multiple chips, each shard is a concatenation of the shards of the individual tensors.
+    For example, let the concatenated_tensor be:
+        AAAAAAAAAAAABBBBBBBBCCCC
+            12 As     8 Bs  4 Cs
+    and let the split_sizes = [12, 8, 4] and n_shards = 4.
+    The output is:
+        AAABBCAAABBCAAABBCAAABBC
+    In other words, it reorders the input tensor into 4 segements, with each segment corresponding to a shard and being AAABBC.
+    Args:
+        concatenated_tensor: the tensor, concatenated on the dimension specified by `dim`.
+        split_sizes: each individual tensor's size on the dimension specified by `dim`.
+        n_shards: num of shards.
+        dim: the dimension on which the concatenated_tensor is concatenated.
+    """
+    # Split the concatenated tensor into individual tensors.
+    split_tensors = []
+    start_offset = 0
+    old_shape = concatenated_tensor.shape
+    # New shape ensures each split_tensor[i] maps to a tensor in ith shards
+    new_shape = old_shape[:dim] + (n_shards, -1) + old_shape[dim + 1:]
+    for split_size in split_sizes:
+        split_tensor = jax.lax.slice_in_dim(concatenated_tensor,
+                                            start_offset,
+                                            start_offset + split_size,
+                                            axis=dim)
+        split_tensors.append(split_tensor.reshape(new_shape))
+        start_offset += split_size
+    # While maintaining 0th dim as a shard dim, we concatenate along 1th dim to
+    # to create concatenated tnensor where 0th dim maps to shard dim.
+    reordered_tensor = jnp.concatenate(split_tensors, axis=dim + 1)
+    return reordered_tensor.reshape(old_shape)
+
+
+def slice_sharded_tensor_for_concatenation(sharded_tensor: jax.Array,
+                                           split_sizes: list[int],
+                                           n_shards: int):
+    """
+    Slice the input tensor which is sharded on multiple chips (on the last dim) into individual tensors with the same sharding.
+    For example, let the sharded_tensor be:
+        AAABBC | AAABBC | AAABBC | AAABBC
+        Shard0   Shard1   Shard2   Shard3
+    and let the split_sizes = [12, 8, 4] and n_shards = 4.
+    The output is a list of 3 tensors:
+         AAA   |  AAA   |  AAA   |  AAA
+          BB   |   BB   |   BB   |   BB
+           C   |    C   |    C   |    C
+        Shard0   Shard1   Shard2   Shard3
+    In other words, each individual tensor is a slice of the input tensor with the same sharding.
+    Args:
+        sharded_tensor: the input tensor, sharded on the last dim.
+        split_sizes: each individual tensor's size on the last dim.
+        n_shards: num of shards.
+    """
+    new_shape = sharded_tensor.shape[:-1] + (n_shards, -1)
+    # New shape ensures each sharded_tensor[:, i] maps to a tensor in ith shards
+    sharded_tensor = sharded_tensor.reshape(new_shape)
+
+    split_tensors = []
+    start_offset = 0
+    for split_size in split_sizes:
+        assert split_size % n_shards == 0
+        sz = split_size // n_shards  # size of this split tensor per shard
+        end_offset = start_offset + sz
+        # Because we are slicing over last dim, sharding dim remains intact.
+        # Therefore, splitting happens locally.
+        split_tensor = sharded_tensor[..., start_offset:end_offset]
+        split_tensors.append(split_tensor.reshape(new_shape[:-2] + (-1, )))
+        start_offset = end_offset
+
+    return split_tensors
+
+
+def torch_to_jax_param(
+    tensor: torch.Tensor,
+    sharding: NamedSharding,
+    output_sizes: Optional[int],
+    n_shards: int,
+    fused: bool,
+    dim: int = 0,
+    jax_dtype: Optional[jnp.dtype] = None,
+) -> Union[torch.nn.Parameter, torch.nn.ParameterList]:
+    if output_sizes is None:
+        output_sizes = [tensor.shape[0]]
+
+    tensor = t2j(tensor, use_dlpack=False)
+    if jax_dtype:
+        tensor = tensor.astype(jax_dtype)
+
+    if fused:
+        tensor = reorder_concatenated_tensor_for_sharding(
+            tensor, output_sizes, n_shards, dim)
+        tensor = jax.device_put(tensor, sharding)
+        param = torch.nn.Parameter(torch_view(tensor), requires_grad=False)
+    else:
+        tensors = []
+        start_offset = 0
+        for size in output_sizes:
+            end_offset = start_offset + size
+
+            tensor_split = jax.lax.slice_in_dim(tensor,
+                                                start_offset,
+                                                end_offset,
+                                                axis=dim)
+            tensor_split = jax.device_put(tensor_split, sharding)
+            tensor_split = torch.nn.Parameter(torch_view(tensor_split),
+                                              requires_grad=False)
+            tensors.append(tensor_split)
+
+            start_offset = end_offset
+        param = torch.nn.ParameterList(tensors)
+    return param
+
+
+MODEL_MATMUL_FUSION_TRUTH_TABLE = {
+    ("Qwen/Qwen2.5-7B-Instruct", 1024, 1, "QKVParallelLinear"):
+    True,
+    ("Qwen/Qwen2.5-7B-Instruct", 1024, 1, "MergedColumnParallelLinear"):
+    False,
+    ("Qwen/Qwen2.5-7B-Instruct", 2048, 1, "QKVParallelLinear"):
+    False,
+    ("Qwen/Qwen2.5-7B-Instruct", 2048, 1, "MergedColumnParallelLinear"):
+    False,
+    ("meta-llama/Llama-3.1-8B-Instruct", 1024, 1, "QKVParallelLinear"):
+    False,
+    ("meta-llama/Llama-3.1-8B-Instruct", 1024, 1, "MergedColumnParallelLinear"):
+    False,
+    ("meta-llama/Llama-3.1-8B-Instruct", 2048, 1, "QKVParallelLinear"):
+    False,
+    ("meta-llama/Llama-3.1-8B-Instruct", 2048, 1, "MergedColumnParallelLinear"):
+    False,
+    ("RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8", 1024, 1, "QKVParallelLinear"):
+    False,
+    ("RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8", 1024, 1, "MergedColumnParallelLinear"):
+    False,
+    ("RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8", 2048, 1, "QKVParallelLinear"):
+    False,
+    ("RedHatAI/Meta-Llama-3.1-8B-Instruct-quantized.w8a8", 2048, 1, "MergedColumnParallelLinear"):
+    False,
+}
+
+
+def get_model_matmul_fusion_assignment(model_name: str, batch_size: int,
+                                       tp_size: int, layer_name: str):
+    key = (model_name, batch_size, tp_size, layer_name)
+    return MODEL_MATMUL_FUSION_TRUTH_TABLE.get(key, True)

tpu_inference/layers/vllm/quantization/__init__.py

@@ -0,0 +1,39 @@
+import copy
+
+from jax.sharding import Mesh
+from vllm.config import VllmConfig
+from vllm.model_executor.layers.quantization.base_config import \
+    QuantizationConfig
+
+from tpu_inference.layers.common import quant_methods
+from tpu_inference.layers.vllm.quantization.awq import VllmAWQConfig
+from tpu_inference.layers.vllm.quantization.common import JaxCommonConfig
+from tpu_inference.layers.vllm.quantization.compressed_tensors.compressed_tensors import \
+    VllmCompressedTensorsConfig  # noqa: E501
+from tpu_inference.layers.vllm.quantization.mxfp4 import VllmMxfp4Config
+from tpu_inference.layers.vllm.quantization.unquantized import \
+    VllmUnquantizedConfig
+
+
+def get_tpu_quantization_config(vllm_config: VllmConfig,
+                                mesh: Mesh) -> QuantizationConfig:
+    model_config = copy.deepcopy(vllm_config.model_config)
+    # TODO(kyuyeunk): Add support for "tpu_int8".
+    method_to_config: dict[str, str] = {
+        None: VllmUnquantizedConfig,
+        quant_methods.COMPRESSED_TENSORS: VllmCompressedTensorsConfig,
+        quant_methods.AWQ: VllmAWQConfig,
+        quant_methods.MXFP4: VllmMxfp4Config,
+    }
+    if model_config.quantization not in method_to_config:
+        raise NotImplementedError(
+            f"{model_config.quantization} quantization method not supported."
+            f" Supported methods are {method_to_config.keys()}")
+    quant_config = method_to_config[model_config.quantization]
+    assert issubclass(quant_config, JaxCommonConfig)
+    quant_config.set_configs(vllm_config, mesh)
+
+    model_config.quantization = quant_methods.get_tpu_quant_method(
+        quant_config.get_name())
+    return VllmConfig.get_quantization_config(model_config,
+                                              vllm_config.load_config)