Improve benchmark scripts & add more models (#484)

benchmark/latency_throughput/bench_throughput.py

@@ -233,8 +233,10 @@ def main(args: argparse.Namespace):
     if args.dataset:
         input_requests = sample_requests(args.dataset, args.num_prompts, tokenizer)
     else:
-        input_lens = np.random.randint(args.input_len * args.range_ratio, args.input_len + 1, size=args.num_prompts)
-        output_lens = np.random.randint(args.output_len * args.range_ratio, args.output_len + 1, size=args.num_prompts)
+        input_lens = np.random.randint(
+            int(args.input_len * args.range_ratio), args.input_len + 1, size=args.num_prompts)
+        output_lens = np.random.randint(
+            int(args.output_len * args.range_ratio), args.output_len + 1, size=args.num_prompts)
         offsets = np.random.randint(0, tokenizer.vocab_size, size=args.num_prompts)
         input_requests = []
         for i in range(args.num_prompts):
@@ -284,13 +286,13 @@ if __name__ == "__main__":
         choices=["vllm", "tgi", "srt", "lightllm", "xinfer"],
     )
     parser.add_argument("--host", type=str, default="localhost")
-    parser.add_argument("--port", type=int, default=8000)
+    parser.add_argument("--port", type=int, default=30000)
     parser.add_argument(
         "--dataset", type=str, help="Path to the dataset."
     )
-    parser.add_argument("--input-len", type=str, default=1024)
-    parser.add_argument("--output-len", type=str, default=128)
-    parser.add_argument("--range-ratio", type=float, default=1.0)
+    parser.add_argument("--input-len", type=str, default=2048)
+    parser.add_argument("--output-len", type=str, default=256)
+    parser.add_argument("--range-ratio", type=float, default=0.5)
     parser.add_argument(
         "--tokenizer", type=str,
         default="NousResearch/Meta-Llama-3-8B",

benchmark/mmlu/bench_other.py

@@ -170,4 +170,4 @@ if __name__ == "__main__":
     parser.add_argument("--data_dir", type=str, default="data")
     parser.add_argument("--nsub", type=int, default=60)
     args = add_common_other_args_and_parse(parser)
-    main(args)
+    main(args)
\ No newline at end of file

python/sglang/srt/layers/fused_moe.py

+# Adapted from
+# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/layers/fused_moe/fused_moe.py#L1
+"""Fused MoE kernel."""
+import functools
+import json
+import os
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from vllm import _custom_ops as ops
+from vllm.logger import init_logger
+from vllm.utils import is_hip
+
+logger = init_logger(__name__)
+
+
+@triton.jit
+def fused_moe_kernel(
+    # Pointers to matrices
+    a_ptr,
+    b_ptr,
+    c_ptr,
+    a_scale_ptr,
+    b_scale_ptr,
+    topk_weights_ptr,
+    sorted_token_ids_ptr,
+    expert_ids_ptr,
+    num_tokens_post_padded_ptr,
+    # Matrix dimensions
+    N,
+    K,
+    EM,
+    num_valid_tokens,
+    # The stride variables represent how much to increase the ptr by when
+    # moving by 1 element in a particular dimension. E.g. `stride_am` is
+    # how much to increase `a_ptr` by to get the element one row down
+    # (A has M rows).
+    stride_am,
+    stride_ak,
+    stride_be,
+    stride_bk,
+    stride_bn,
+    stride_cm,
+    stride_cn,
+    # Meta-parameters
+    BLOCK_SIZE_M: tl.constexpr,
+    BLOCK_SIZE_N: tl.constexpr,
+    BLOCK_SIZE_K: tl.constexpr,
+    GROUP_SIZE_M: tl.constexpr,
+    MUL_ROUTED_WEIGHT: tl.constexpr,
+    top_k: tl.constexpr,
+    compute_type: tl.constexpr,
+    use_fp8: tl.constexpr,
+):
+    """
+    Implements the fused computation for a Mixture of Experts (MOE) using
+    token and expert matrices.
+
+    Key Parameters:
+    - A: The input tensor representing tokens with shape (*, K), where '*' can
+        be any shape representing batches and K is the feature dimension of
+        each token.
+    - B: The stacked MOE weight tensor with shape (E, N, K), where E is
+        the number of experts, K is the input feature dimension, and N is
+        the output feature dimension.
+    - C: The output cache tensor with shape (M, topk, N), where M is the
+        total number of tokens post padding, topk is the number of times
+        each token is repeated, and N is the output feature dimension.
+    - sorted_token_ids: A tensor containing the sorted indices of tokens,
+        repeated topk times and arranged by the expert index they are
+        assigned to.
+    - expert_ids: A tensor containing the indices of the expert for each
+        block. It determines which expert matrix from B should be used for
+        each block in A.
+    This kernel performs the multiplication of a token by its corresponding
+    expert matrix as determined by `expert_ids`. The sorting of
+    `sorted_token_ids` by expert index and padding ensures divisibility by
+    BLOCK_SIZE_M, which is necessary to maintain consistency in block matrix
+    multiplication across different blocks processed by the same expert.
+    """
+    # -----------------------------------------------------------
+    # Map program ids `pid` to the block of C it should compute.
+    # This is done in a grouped ordering to promote L2 data reuse.
+    pid = tl.program_id(axis=0)
+    num_pid_m = tl.cdiv(EM, BLOCK_SIZE_M)
+    num_pid_n = tl.cdiv(N, BLOCK_SIZE_N)
+    num_pid_in_group = GROUP_SIZE_M * num_pid_n
+    group_id = pid // num_pid_in_group
+    first_pid_m = group_id * GROUP_SIZE_M
+    group_size_m = min(num_pid_m - first_pid_m, GROUP_SIZE_M)
+    pid_m = first_pid_m + ((pid % num_pid_in_group) % group_size_m)
+    pid_n = (pid % num_pid_in_group) // group_size_m
+
+    # ----------------------------------------------------------
+    # Create pointers for the first blocks of A and B.
+    # We will advance this pointer as we move in the K direction
+    # and accumulate
+    # `a_ptrs` is a block of [BLOCK_SIZE_M, BLOCK_SIZE_K] pointers
+    # `b_ptrs` is a block of [BLOCK_SIZE_K, BLOCK_SIZE_N] pointers
+    num_tokens_post_padded = tl.load(num_tokens_post_padded_ptr)
+    if pid_m * BLOCK_SIZE_M >= num_tokens_post_padded:
+        return
+    offs_token_id = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
+    offs_token = tl.load(sorted_token_ids_ptr + offs_token_id)
+    token_mask = offs_token < num_valid_tokens
+
+    offs_bn = (pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)) % N
+    offs_k = tl.arange(0, BLOCK_SIZE_K)
+    a_ptrs = a_ptr + (offs_token[:, None] // top_k * stride_am +
+                      offs_k[None, :] * stride_ak)
+
+    off_experts = tl.load(expert_ids_ptr + pid_m)
+    b_ptrs = b_ptr + off_experts * stride_be + (offs_k[:, None] * stride_bk +
+                                                offs_bn[None, :] * stride_bn)
+
+    if use_fp8:
+        a_scale = tl.load(a_scale_ptr)
+        b_scale = tl.load(b_scale_ptr + off_experts)
+
+    # -----------------------------------------------------------
+    # Iterate to compute a block of the C matrix.
+    # We accumulate into a `[BLOCK_SIZE_M, BLOCK_SIZE_N]` block
+    # of fp32 values for higher accuracy.
+    # `accumulator` will be converted back to fp16 after the loop.
+    accumulator = tl.zeros((BLOCK_SIZE_M, BLOCK_SIZE_N), dtype=tl.float32)
+
+    for k in range(0, tl.cdiv(K, BLOCK_SIZE_K)):
+        # Load the next block of A and B, generate a mask by checking the
+        # K dimension.
+        a = tl.load(a_ptrs,
+                    mask=token_mask[:, None] &
+                    (offs_k[None, :] < K - k * BLOCK_SIZE_K),
+                    other=0.0)
+        b = tl.load(b_ptrs,
+                    mask=offs_k[:, None] < K - k * BLOCK_SIZE_K,
+                    other=0.0)
+        # We accumulate along the K dimension.
+        if use_fp8:
+            accumulator = tl.dot(a, b, acc=accumulator)
+        else:
+            accumulator += tl.dot(a, b)
+        # Advance the ptrs to the next K block.
+        a_ptrs += BLOCK_SIZE_K * stride_ak
+        b_ptrs += BLOCK_SIZE_K * stride_bk
+
+    if MUL_ROUTED_WEIGHT:
+        moe_weight = tl.load(topk_weights_ptr + offs_token,
+                             mask=token_mask,
+                             other=0)
+        accumulator = accumulator * moe_weight[:, None]
+
+    if use_fp8:
+        accumulator = (accumulator * a_scale * b_scale).to(compute_type)
+    else:
+        accumulator = accumulator.to(compute_type)
+    # -----------------------------------------------------------
+    # Write back the block of the output
+    offs_cn = pid_n * BLOCK_SIZE_N + tl.arange(0, BLOCK_SIZE_N)
+    c_ptrs = c_ptr + stride_cm * offs_token[:, None] + stride_cn * offs_cn[
+        None, :]
+    c_mask = token_mask[:, None] & (offs_cn[None, :] < N)
+    tl.store(c_ptrs, accumulator, mask=c_mask)
+
+
+def moe_align_block_size(
+        topk_ids: torch.Tensor, block_size: int,
+        num_experts: int) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """
+    Aligns the token distribution across experts to be compatible with block
+    size for matrix multiplication.
+
+    Parameters:
+    - topk_ids: A tensor of shape [total_tokens, top_k] representing the
+        top-k expert indices for each token.
+    - block_size: The block size used in block matrix multiplication.
+    - num_experts: The total number of experts.
+
+    Returns:
+    - sorted_token_ids: A tensor containing the sorted token indices according
+        to their allocated expert.
+    - expert_ids: A tensor indicating the assigned expert index for each block.
+    - num_tokens_post_padded: The total number of tokens after padding,
+        ensuring divisibility by block_size.
+
+    This function pads the number of tokens that each expert needs to process
+    so that it is divisible by block_size.
+    Padding ensures that during block matrix multiplication, the dimensions
+    align correctly.
+
+    Example:
+    Given topk_ids = [[2, 3, 4], [1, 2, 4], [1, 3, 4], [1, 2, 3]],
+    block_size = 4, and num_experts = 4:
+    - We initially have 12 tokens (after repeating 'top_k' times) and 4 experts,
+        with each expert needing to process 3 tokens.
+    - As block_size is 4, we pad 1 token for each expert.
+    - First, flatten topk_ids to [2, 3, 4, 1, 2, 4, 1, 3, 4, 1, 2, 3].
+    - Then append padding tokens [12, 12, 12, 12] for each block.
+    - After sorting by expert index, we obtain token_ids
+        [3, 6, 9, 12, 0, 4, 10, 12, 1, 7, 11, 12, 2, 5, 8, 12].
+        Tokens 12 are non-existent (padding) and are ignored in
+        the subsequent matrix multiplication.
+    - The padding ensures that the total number of tokens is now divisible
+        by block_size for proper block matrix operations.
+    """
+    max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1)
+    sorted_ids = torch.empty((max_num_tokens_padded, ),
+                             dtype=torch.int32,
+                             device=topk_ids.device)
+    sorted_ids.fill_(topk_ids.numel())
+    max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size)
+    expert_ids = torch.empty((max_num_m_blocks, ),
+                             dtype=torch.int32,
+                             device=topk_ids.device)
+    num_tokens_post_pad = torch.empty((1),
+                                      dtype=torch.int32,
+                                      device=topk_ids.device)
+    ops.moe_align_block_size(topk_ids, num_experts, block_size, sorted_ids,
+                             expert_ids, num_tokens_post_pad)
+    return sorted_ids, expert_ids, num_tokens_post_pad
+
+
+def invoke_fused_moe_kernel(A: torch.Tensor, B: torch.Tensor, C: torch.Tensor,
+                            A_scale: Optional[torch.Tensor],
+                            B_scale: Optional[torch.Tensor],
+                            topk_weights: torch.Tensor, topk_ids: torch.Tensor,
+                            sorted_token_ids: torch.Tensor,
+                            expert_ids: torch.Tensor,
+                            num_tokens_post_padded: torch.Tensor,
+                            mul_routed_weight: bool, top_k: int,
+                            config: Dict[str, Any], compute_type: tl.dtype,
+                            use_fp8: bool) -> None:
+    assert topk_weights.stride(1) == 1
+    assert sorted_token_ids.stride(0) == 1
+
+    if not use_fp8:
+        assert A_scale is None
+        assert B_scale is None
+    else:
+        A, A_scale = ops.scaled_fp8_quant(A, A_scale)
+        assert B_scale is not None
+
+    grid = lambda META: (triton.cdiv(sorted_token_ids.shape[0], META[
+        'BLOCK_SIZE_M']) * triton.cdiv(B.shape[1], META['BLOCK_SIZE_N']), )
+
+    fused_moe_kernel[grid](
+        A,
+        B,
+        C,
+        A_scale,
+        B_scale,
+        topk_weights,
+        sorted_token_ids,
+        expert_ids,
+        num_tokens_post_padded,
+        B.shape[1],
+        B.shape[2],
+        sorted_token_ids.shape[0],
+        topk_ids.numel(),
+        A.stride(0),
+        A.stride(1),
+        B.stride(0),
+        B.stride(2),
+        B.stride(1),
+        C.stride(1),
+        C.stride(2),
+        MUL_ROUTED_WEIGHT=mul_routed_weight,
+        top_k=top_k,
+        compute_type=compute_type,
+        use_fp8=use_fp8,
+        **config,
+    )
+
+
+def get_config_file_name(E: int, N: int, dtype: Optional[str]) -> str:
+    device_name = torch.cuda.get_device_name().replace(" ", "_")
+    dtype_selector = "" if not dtype else f",dtype={dtype}"
+    return f"E={E},N={N},device_name={device_name}{dtype_selector}.json"
+
+
+@functools.lru_cache
+def get_moe_configs(E: int, N: int,
+                    dtype: Optional[str]) -> Optional[Dict[int, Any]]:
+    """
+    Return optimized configurations for the fused MoE kernel.
+
+    The return value will be a dictionary that maps an irregular grid of
+    batch sizes to configurations of the fused_moe kernel. To evaluate the
+    kernel on a given batch size bs, the closest batch size in the grid should
+    be picked and the associated configuration chosen to invoke the kernel.
+    """
+
+    # First look up if an optimized configuration is available in the configs
+    # directory
+    json_file_name = get_config_file_name(E, N, dtype)
+
+    config_file_path = os.path.join(
+        os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
+    if os.path.exists(config_file_path):
+        with open(config_file_path) as f:
+            logger.info("Using configuration from %s for MoE layer.",
+                        config_file_path)
+            # If a configuration has been found, return it
+            return {int(key): val for key, val in json.load(f).items()}
+
+    # If no optimized configuration is available, we will use the default
+    # configuration
+    return None
+
+
+def fused_moe(
+    hidden_states: torch.Tensor,
+    w1: torch.Tensor,
+    w2: torch.Tensor,
+    gating_output: torch.Tensor,
+    topk: int,
+    renormalize: bool,
+    inplace: bool = False,
+    override_config: Optional[Dict[str, Any]] = None,
+    use_fp8: bool = False,
+    w1_scale: Optional[torch.Tensor] = None,
+    w2_scale: Optional[torch.Tensor] = None,
+    a1_scale: Optional[torch.Tensor] = None,
+    a2_scale: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """
+    This function computes a Mixture of Experts (MoE) layer using two sets of
+    weights, w1 and w2, and top-k gating mechanism.
+
+    Parameters:
+    - hidden_states (torch.Tensor): The input tensor to the MoE layer.
+    - w1 (torch.Tensor): The first set of expert weights.
+    - w2 (torch.Tensor): The second set of expert weights.
+    - gating_output (torch.Tensor): The output of the gating operation
+        (before softmax).
+    - topk (int): The number of top-k experts to select.
+    - renormalize (bool): If True, renormalize the top-k weights to sum to 1.
+    - inplace (bool): If True, perform the operation in-place.
+        Defaults to False.
+    - override_config (Optional[Dict[str, Any]]): Optional override
+        for the kernel configuration.
+    - use_fp8 (bool): If True, use fp8 arithmetic to compute the inner
+        products for w1 and w2. Defaults to False.
+    - w1_scale (Optional[torch.Tensor]): Optional scale to be used for
+        w1.
+    - w2_scale (Optional[torch.Tensor]): Optional scale to be used for
+        w2.
+
+    Returns:
+    - torch.Tensor: The output tensor after applying the MoE layer.
+    """
+    # Check constraints.
+    assert hidden_states.shape[0] == gating_output.shape[0], (
+        "Number of tokens mismatch")
+    assert hidden_states.shape[1] == w1.shape[2], "Hidden size mismatch"
+    assert gating_output.shape[1] == w1.shape[0], "Number of experts mismatch"
+    assert hidden_states.is_contiguous(), "Hidden_states must be contiguous"
+    assert w1.is_contiguous(), "Expert weights1 must be contiguous"
+    assert w2.is_contiguous(), "Expert weights2 must be contiguous"
+    assert hidden_states.dtype in [
+        torch.float32, torch.float16, torch.bfloat16
+    ]
+    M, _ = hidden_states.shape
+    E, N, _ = w1.shape
+
+    if is_hip():
+        # The MoE kernels are not yet supported on ROCm.
+        routing_weights = torch.softmax(gating_output,
+                                        dim=-1,
+                                        dtype=torch.float32)
+        topk_weights, topk_ids = torch.topk(routing_weights, topk, dim=-1)
+    else:
+        import vllm._moe_C as moe_kernels
+
+        topk_weights = torch.empty(M,
+                                   topk,
+                                   dtype=torch.float32,
+                                   device=hidden_states.device)
+        topk_ids = torch.empty(M,
+                               topk,
+                               dtype=torch.int32,
+                               device=hidden_states.device)
+        token_expert_indicies = torch.empty(M,
+                                            topk,
+                                            dtype=torch.int32,
+                                            device=hidden_states.device)
+        moe_kernels.topk_softmax(
+            topk_weights,
+            topk_ids,
+            token_expert_indicies,
+            gating_output.float(),  # TODO(woosuk): Optimize this.
+        )
+        del token_expert_indicies  # Not used. Will be used in the future.
+    if renormalize:
+        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)
+
+    if override_config:
+        config = override_config
+    else:
+        # First try to load optimal config from the file
+        configs = get_moe_configs(E, w2.shape[2],
+                                  "float8" if use_fp8 else None)
+
+        if configs:
+            # If an optimal configuration map has been found, look up the
+            # optimal config
+            config = configs[min(configs.keys(), key=lambda x: abs(x - M))]
+        else:
+            # Else use the default config
+            config = {
+                "BLOCK_SIZE_M": 128,
+                "BLOCK_SIZE_N": 64,
+                "BLOCK_SIZE_K": 128,
+                "GROUP_SIZE_M": 1,
+                "num_warps": 4,
+                "num_stages": 4
+            }
+
+            if M <= E:
+                config = {
+                    "BLOCK_SIZE_M": 128,
+                    "BLOCK_SIZE_N": 256,
+                    "BLOCK_SIZE_K": 128,
+                    "GROUP_SIZE_M": 16,
+                    "num_warps": 8,
+                    "num_stages": 4
+                }
+
+    intermediate_cache1 = torch.empty((M, topk_ids.shape[1], N),
+                                      device=hidden_states.device,
+                                      dtype=hidden_states.dtype)
+    intermediate_cache2 = torch.empty((M * topk_ids.shape[1], N // 2),
+                                      device=hidden_states.device,
+                                      dtype=hidden_states.dtype)
+    intermediate_cache3 = torch.empty((M, topk_ids.shape[1], w2.shape[1]),
+                                      device=hidden_states.device,
+                                      dtype=hidden_states.dtype)
+
+    sorted_token_ids, expert_ids, num_tokens_post_padded = moe_align_block_size(
+        topk_ids, config['BLOCK_SIZE_M'], E)
+    compute_type = (tl.bfloat16
+                    if hidden_states.dtype == torch.bfloat16 else tl.float16)
+
+    invoke_fused_moe_kernel(hidden_states,
+                            w1,
+                            intermediate_cache1,
+                            a1_scale,
+                            w1_scale,
+                            topk_weights,
+                            topk_ids,
+                            sorted_token_ids,
+                            expert_ids,
+                            num_tokens_post_padded,
+                            False,
+                            topk_ids.shape[1],
+                            config,
+                            compute_type=compute_type,
+                            use_fp8=use_fp8)
+
+    ops.gelu_and_mul(intermediate_cache2, intermediate_cache1.view(-1, N))
+
+    invoke_fused_moe_kernel(intermediate_cache2,
+                            w2,
+                            intermediate_cache3,
+                            a2_scale,
+                            w2_scale,
+                            topk_weights,
+                            topk_ids,
+                            sorted_token_ids,
+                            expert_ids,
+                            num_tokens_post_padded,
+                            True,
+                            1,
+                            config,
+                            compute_type=compute_type,
+                            use_fp8=use_fp8)
+
+    if inplace:
+        return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape),
+                         dim=1,
+                         out=hidden_states)
+    return torch.sum(intermediate_cache3.view(*intermediate_cache3.shape),
+                     dim=1)
\ No newline at end of file

python/sglang/srt/models/grok.py

+# Adapted from
+# https://github.com/vllm-project/vllm/blob/c7f2cf2b7f67bce5842fedfdba508440fe257375/vllm/model_executor/models/mixtral.py#L1
+"""Inference-only Grok1 model."""
+from typing import Iterable, Optional, Tuple, List
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+import tqdm
+from torch import nn
+from transformers import PretrainedConfig
+
+from vllm import _custom_ops as ops
+from vllm.distributed import (
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+    tensor_model_parallel_all_reduce,
+)
+from vllm.model_executor.layers.layernorm import RMSNorm
+from vllm.model_executor.layers.linear import (
+    QKVParallelLinear,
+    ReplicatedLinear,
+    RowParallelLinear,
+)
+from vllm.model_executor.layers.quantization.base_config import QuantizationConfig
+from vllm.model_executor.layers.quantization.fp8 import Fp8Config
+from vllm.model_executor.layers.rotary_embedding import get_rope
+from vllm.model_executor.layers.vocab_parallel_embedding import (
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
+from vllm.model_executor.model_loader.loader import DefaultModelLoader
+from vllm.model_executor.model_loader.weight_utils import default_weight_loader
+from vllm.model_executor.utils import set_weight_attrs
+from vllm.utils import print_warning_once
+
+from sglang.srt.layers.logits_processor import LogitsProcessor
+from sglang.srt.layers.fused_moe import fused_moe
+from sglang.srt.layers.radix_attention import RadixAttention
+from sglang.srt.managers.router.model_runner import InputMetadata
+
+
+use_fused = True
+
+
+class Grok1MLP(nn.Module):
+    def __init__(
+        self,
+        num_experts: int,
+        hidden_size: int,
+        intermediate_size: int,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.num_experts = num_experts
+        self.ffn_dim = intermediate_size
+        self.hidden_dim = hidden_size
+
+        self.w1 = ReplicatedLinear(
+            self.hidden_dim, self.ffn_dim, bias=False, quant_config=quant_config
+        )
+        self.w2 = ReplicatedLinear(
+            self.ffn_dim, self.hidden_dim, bias=False, quant_config=quant_config
+        )
+        self.w3 = ReplicatedLinear(
+            self.hidden_dim, self.ffn_dim, bias=False, quant_config=quant_config
+        )
+
+        self.act_fn = nn.GELU()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        w1_out, _ = self.w1(hidden_states)
+        w1_out = self.act_fn(w1_out)
+        w3_out, _ = self.w3(hidden_states)
+        current_hidden_states = w1_out * w3_out
+        current_hidden_states, _ = self.w2(current_hidden_states)
+        return current_hidden_states
+
+
+class Grok1MoEUnfused(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.config = config
+        self.rank = get_tensor_model_parallel_rank()
+        self.tp_size = get_tensor_model_parallel_world_size()
+        self.num_total_experts = config.num_local_experts
+        self.top_k = config.num_experts_per_tok
+        if self.tp_size > self.num_total_experts:
+            raise ValueError(
+                f"Tensor parallel size {self.tp_size} is greater than "
+                f"the number of experts {self.num_total_experts}."
+            )
+        # Split experts equally between ranks
+        self.expert_indicies = np.array_split(
+            range(self.num_total_experts), self.tp_size
+        )[self.rank].tolist()
+        if not self.expert_indicies:
+            raise ValueError(f"Rank {self.rank} has no experts assigned to it.")
+
+        self.experts = nn.ModuleList(
+            [
+                (
+                    Grok1MLP(
+                        self.num_total_experts,
+                        config.hidden_size,
+                        config.intermediate_size,
+                        quant_config=quant_config,
+                    )
+                    if idx in self.expert_indicies
+                    else None
+                )
+                for idx in range(self.num_total_experts)
+            ]
+        )
+        self.gate = ReplicatedLinear(
+            config.hidden_size, self.num_total_experts, bias=False, quant_config=None
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        router_logits, _ = self.gate(hidden_states)
+        router_logits = 30 * F.tanh(router_logits / 30)
+
+        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(
+            routing_weights, self.top_k, dim=-1
+        )
+        routing_weights = routing_weights.to(hidden_states.dtype)
+        hidden_dim = hidden_states.shape[1]
+
+        final_hidden_states = torch.zeros(
+            (hidden_states.shape[0], hidden_dim),
+            dtype=hidden_states.dtype, device=hidden_states.device
+        )
+        expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_total_experts).permute(2, 1, 0)
+
+        for expert_idx in self.expert_indicies:
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            if top_x.shape[0] == 0:
+                continue
+
+            # in torch it is faster to index using lists than torch tensors
+            top_x_list = top_x.tolist()
+            idx_list = idx.tolist()
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x_list].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x_list, idx_list, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states)
+
+        return tensor_model_parallel_all_reduce(final_hidden_states)
+
+
+class Grok1MoE(nn.Module):
+    """A tensor-parallel MoE implementation for Grok1 that shards each expert
+    across all ranks.
+
+    Each expert's weights are sharded across all ranks and a fused MoE
+    kernel is used for the forward pass, and finally we reduce the outputs
+    across ranks.
+    """
+
+    def __init__(
+        self,
+        num_experts: int,
+        top_k: int,
+        hidden_size: int,
+        intermediate_size: int,
+        params_dtype: Optional[torch.dtype] = None,
+        tp_size: Optional[int] = None,
+        quant_config: Optional[QuantizationConfig] = None,
+    ):
+        super().__init__()
+        self.tp_size = tp_size or get_tensor_model_parallel_world_size()
+        self.num_total_experts = num_experts
+        self.top_k = top_k
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size // self.tp_size
+        self.quant_config = quant_config
+
+        # FIXME(pcmoritz): Make this more general to support different
+        # quantization schemes
+        self.use_fp8 = isinstance(quant_config, Fp8Config)
+
+        if params_dtype is None:
+            params_dtype = torch.get_default_dtype()
+        self.params_dtype = params_dtype
+
+        # Gate always runs at half / full precision for now.
+        self.gate = ReplicatedLinear(self.hidden_size,
+                                     self.num_total_experts,
+                                     bias=False,
+                                     params_dtype=self.params_dtype,
+                                     quant_config=None)
+
+        if self.use_fp8 and self.quant_config.is_checkpoint_fp8_serialized:
+            params_dtype = torch.float8_e4m3fn
+
+        self.w13_weight = nn.Parameter(
+            torch.empty(self.num_total_experts,
+                        2 * self.intermediate_size,
+                        self.hidden_size,
+                        dtype=params_dtype))
+        self.w2_weight = nn.Parameter(
+            torch.empty(self.num_total_experts,
+                        self.hidden_size,
+                        self.intermediate_size,
+                        dtype=params_dtype))
+
+        set_weight_attrs(self.w13_weight, {
+            "weight_loader": self.weight_loader,
+        })
+        set_weight_attrs(self.w2_weight, {
+            "weight_loader": self.weight_loader,
+        })
+
+        # Used for fp8.
+        self.w13_scale = None
+        self.w2_scale = None
+        self.a13_scale = None
+        self.a2_scale = None
+
+        if self.use_fp8:
+            # WEIGHT_SCALE (for fp8)
+            self.w13_scale = nn.Parameter(torch.ones(self.num_total_experts,
+                                                     dtype=torch.float32),
+                                          requires_grad=False)
+            self.w2_scale = nn.Parameter(torch.ones(self.num_total_experts,
+                                                    dtype=torch.float32),
+                                         requires_grad=False)
+
+            # If loading fp8 checkpoint, pass the weight loaders.
+            # If loading an fp16 checkpoint, do not (we will quantize in
+            #   process_weights_after_loading()
+            if quant_config.is_checkpoint_fp8_serialized:
+                set_weight_attrs(self.w13_scale, {
+                    "weight_loader": self.weight_loader,
+                })
+                set_weight_attrs(self.w2_scale, {
+                    "weight_loader": self.weight_loader,
+                })
+
+            # ACT_SCALE (for fp8)
+            if quant_config.activation_scheme == "static":
+                if not quant_config.is_checkpoint_fp8_serialized:
+                    raise ValueError(
+                        "Found static activation scheme for checkpoint that "
+                        "was not serialized fp8.")
+                self.a13_scale = nn.Parameter(torch.zeros(
+                    self.num_total_experts, dtype=torch.float32),
+                                              requires_grad=False)
+                self.a2_scale = nn.Parameter(torch.zeros(
+                    self.num_total_experts, dtype=torch.float32),
+                                             requires_grad=False)
+
+                set_weight_attrs(self.a13_scale, {
+                    "weight_loader": self.weight_loader,
+                })
+                set_weight_attrs(self.a2_scale, {
+                    "weight_loader": self.weight_loader,
+                })
+
+    def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor,
+                      weight_name: str, expert_id: int, pre_sharded: bool):
+        param_data = param.data
+        shard_size = self.intermediate_size
+        if pre_sharded:
+            # The weight is already sharded. Readl the full shard
+            shard = slice(None)
+        else:
+            tp_rank = get_tensor_model_parallel_rank()
+            shard = slice(tp_rank * shard_size, (tp_rank + 1) * shard_size)
+        if weight_name.endswith("w1.weight"):
+            param_data[expert_id, 0:shard_size, :] = loaded_weight[shard, :]
+        if weight_name.endswith("w3.weight"):
+            param_data[expert_id,
+                       shard_size:2 * shard_size, :] = loaded_weight[shard, :]
+        if weight_name.endswith("w2.weight"):
+            param_data[expert_id, :, :] = loaded_weight[:, shard]
+        if "act_scale" in weight_name or "weight_scale" in weight_name:
+            param_data[expert_id] = loaded_weight
+
+    def process_weights_after_loading(self):
+        # Fp8 is the only case where we need to process after loading.
+        if not self.use_fp8:
+            return
+
+        # If checkpoint is fp16, quantize here.
+        if not self.quant_config.is_checkpoint_fp8_serialized:
+            w13_weight = torch.empty_like(self.w13_weight.data,
+                                          dtype=torch.float8_e4m3fn)
+            w2_weight = torch.empty_like(self.w2_weight.data,
+                                         dtype=torch.float8_e4m3fn)
+            for expert in range(self.num_total_experts):
+                w13_weight[expert, :, :], self.w13_scale[
+                    expert] = ops.scaled_fp8_quant(
+                        self.w13_weight.data[expert, :, :])
+                w2_weight[expert, :, :], self.w2_scale[
+                    expert] = ops.scaled_fp8_quant(
+                        self.w2_weight.data[expert, :, :])
+            self.w13_weight = nn.Parameter(w13_weight, requires_grad=False)
+            self.w2_weight = nn.Parameter(w2_weight, requires_grad=False)
+
+        # If checkpoint is fp8 + static, cleanup act_scales.
+        #   Since state_dict has an act_scale per expert but our kernels
+        #   are passed one act_scale shared across all experts.
+        elif self.quant_config.activation_scheme == "static":
+            if self.a13_scale is None or self.a2_scale is None:
+                raise ValueError(
+                    "QuantConfig has static quantization, but found "
+                    "activation scales are None.")
+
+            if (not all_close_1d(self.a13_scale)
+                    or not all_close_1d(self.a2_scale)):
+                print_warning_once(
+                    "Found act_scales that are not equal for fp8 MoE layer. "
+                    "Using the maximum across experts for each layer. ")
+
+            self.a13_scale = nn.Parameter(self.a13_scale.max(),
+                                          requires_grad=False)
+            self.a2_scale = nn.Parameter(self.a2_scale.max(),
+                                         requires_grad=False)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        num_tokens, hidden_size = hidden_states.shape
+        hidden_states = hidden_states.view(-1, self.hidden_size)
+        # router_logits: (num_tokens, n_experts)
+        router_logits, _ = self.gate(hidden_states)
+        final_hidden_states = fused_moe(hidden_states,
+                                        self.w13_weight,
+                                        self.w2_weight,
+                                        router_logits,
+                                        self.top_k,
+                                        renormalize=False,
+                                        inplace=True,
+                                        use_fp8=self.use_fp8,
+                                        w1_scale=self.w13_scale,
+                                        w2_scale=self.w2_scale,
+                                        a1_scale=self.a13_scale,
+                                        a2_scale=self.a2_scale)
+
+        if self.tp_size > 1:
+            final_hidden_states = tensor_model_parallel_all_reduce(
+                final_hidden_states)
+
+        return final_hidden_states.view(num_tokens, hidden_size)
+
+
+class Grok1Attention(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        layer_id: int = 0,
+        max_position: int = 4096 * 32,
+        rope_theta: float = 10000,
+        logit_cap: float = 30,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        self.head_dim = 128
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=False,
+            quant_config=quant_config,
+        )
+
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+        )
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position,
+            base=int(self.rope_theta),
+            is_neox_style=True,
+        )
+        self.attn = RadixAttention(
+            self.num_heads,
+            self.head_dim,
+            self.scaling,
+            num_kv_heads=self.num_kv_heads,
+            layer_id=layer_id,
+            logit_cap=logit_cap,
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+        qkv, _ = self.qkv_proj(hidden_states)
+        q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
+        q, k = self.rotary_emb(positions, q, k)
+        attn_output = self.attn(q, k, v, input_metadata)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Grok1DecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        # Requires transformers > 4.32.0
+        rope_theta = getattr(config, "rope_theta", 10000)
+        self.self_attn = Grok1Attention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            max_position=config.max_position_embeddings,
+            num_kv_heads=config.num_key_value_heads,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            quant_config=quant_config,
+        )
+        if use_fused:
+            self.block_sparse_moe = Grok1MoE(
+                num_experts=config.num_local_experts,
+                top_k=config.num_experts_per_tok,
+                hidden_size=config.hidden_size,
+                intermediate_size=config.intermediate_size,
+                quant_config=quant_config)
+        else:
+            self.block_sparse_moe = Grok1MoEUnfused(
+                config=config, quant_config=quant_config)        
+        self.pre_attn_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attn_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.pre_moe_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_moe_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        input_metadata: InputMetadata,
+    ) -> torch.Tensor:
+
+        hidden_states = self.post_attn_norm(self.self_attn(
+            positions=positions, hidden_states=self.pre_attn_norm(hidden_states),
+            input_metadata=input_metadata,
+        )) + hidden_states
+
+        hidden_states = self.post_moe_norm(self.block_sparse_moe(self.pre_moe_norm(hidden_states))) + hidden_states
+
+        return hidden_states
+
+
+class Grok1Model(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+
+        self.embed_tokens = VocabParallelEmbedding(
+            config.vocab_size,
+            config.hidden_size,
+        )
+        self.layers = nn.ModuleList(
+            [
+                Grok1DecoderLayer(config, i, quant_config=quant_config)
+                for i in range(config.num_hidden_layers)
+            ]
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        if input_embeds is None:
+            hidden_states = self.embed_tokens(input_ids)
+        else:
+            hidden_states = input_embeds
+        hidden_states.mul_(self.config.embedding_multiplier_scale)
+
+        for i in range(len(self.layers)):
+            hidden_states = self.layers[i](
+                positions, hidden_states, input_metadata
+            )
+
+        hidden_states = self.norm(hidden_states)
+        hidden_states.mul_(self.config.output_multiplier_scale)
+        return hidden_states
+
+
+class Grok1ModelForCausalLM(nn.Module):
+    def __init__(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.quant_config = quant_config
+        self.model = Grok1Model(config, quant_config=quant_config)
+        self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size)
+        self.logits_processor = LogitsProcessor(config)
+
+        # Monkey patch _prepare_weights to load pre-sharded weights
+        setattr(DefaultModelLoader, "_prepare_weights", _prepare_presharded_weights)
+
+    def forward(
+        self,
+        input_ids: torch.Tensor,
+        positions: torch.Tensor,
+        input_metadata: InputMetadata,
+        input_embeds: torch.Tensor = None,
+    ) -> torch.Tensor:
+        hidden_states = self.model(input_ids, positions, input_metadata, input_embeds)
+        return self.logits_processor(
+            input_ids, hidden_states, self.lm_head.weight, input_metadata
+        )
+
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        stacked_params_mapping = [
+            # (param_name, shard_name, shard_id)
+            ("qkv_proj", "q_proj", "q"),
+            ("qkv_proj", "k_proj", "k"),
+            ("qkv_proj", "v_proj", "v"),
+        ]
+
+        if use_fused:
+            expert_params_mapping = [
+                # These are the weight scales for the experts
+                # (param_name, weight_name, expert_id)
+                ("w13_scale" if weight_name in ["w1", "w3"] else "w2_scale",
+                f"experts.{expert_id}.{weight_name}.weight_scale", expert_id)
+                for expert_id in range(self.config.num_local_experts)
+                for weight_name in ["w1", "w2", "w3"]
+            ] + [
+                # These are the weights for the experts
+                # (param_name, weight_name, expert_id)
+                ("w13_weight" if weight_name in ["w1", "w3"] else "w2_weight",
+                f"experts.{expert_id}.{weight_name}.weight", expert_id)
+                for expert_id in range(self.config.num_local_experts)
+                for weight_name in ["w1", "w2", "w3"]
+            ] + [
+                # These are the activation scales for the experts
+                # (param_name, weight_name, expert_id)
+                ("a13_scale" if weight_name in ["w1", "w3"] else "a2_scale",
+                f"experts.{expert_id}.{weight_name}.act_scale", expert_id)
+                for expert_id in range(self.config.num_local_experts)
+                for weight_name in ["w1", "w2", "w3"]
+            ]
+        else:
+            expert_params_mapping = []
+
+        params_dict = dict(self.named_parameters())
+        if get_tensor_model_parallel_rank() == 0:
+            weights = tqdm.tqdm(weights, total=int(len(params_dict) * 3.4))
+        for name, loaded_weight in weights:
+            #print(get_tensor_model_parallel_rank(), name)
+            if "rotary_emb.inv_freq" in name:
+                continue
+
+            for (param_name, weight_name, shard_id) in stacked_params_mapping:
+                if weight_name not in name:
+                    continue
+                name = name.replace(weight_name, param_name)
+                # Skip loading extra bias for GPTQ models.
+                if name.endswith(".bias") and name not in params_dict:
+                    continue
+                param = params_dict[name]
+                weight_loader = param.weight_loader
+                weight_loader(param, loaded_weight, shard_id)
+                break
+            else:
+                for param_name, weight_name, expert_id in expert_params_mapping:
+                    if weight_name not in name:
+                        continue
+                    name = name.replace(weight_name, param_name)
+                    param = params_dict[name]
+                    weight_loader = param.weight_loader
+                    weight_loader(param,
+                                  loaded_weight,
+                                  weight_name,
+                                  expert_id=expert_id,
+                                  pre_sharded=get_tensor_model_parallel_world_size() > 1)
+                    break
+                else:
+                    # Skip loading extra bias for GPTQ models.
+                    if name.endswith(".bias") and name not in params_dict:
+                        continue
+                    param = params_dict[name]
+                    weight_loader = getattr(param, "weight_loader",
+                                            default_weight_loader)
+                    weight_loader(param, loaded_weight)
+
+
+def all_close_1d(x: torch.Tensor) -> bool:
+    assert len(x.shape) == 1
+    return all(torch.allclose(x[0], x[i]) for i in range(x.shape[0]))
+
+
+old_prepare_weights = getattr(DefaultModelLoader, "_prepare_weights")
+def _prepare_presharded_weights(self,
+                                model_name_or_path: str,
+                                revision: Optional[str],
+                                fall_back_to_pt: bool) -> Tuple[str, List[str], bool]:
+    import glob
+    import os
+
+    if get_tensor_model_parallel_world_size() == 1:
+        return old_prepare_weights(self, model_name_or_path, revision, fall_back_to_pt)
+
+    tp_rank = get_tensor_model_parallel_rank()
+    allow_patterns = [f"*-{tp_rank:03d}.bin"]
+
+    hf_folder = model_name_or_path
+
+    hf_weights_files: List[str] = []
+    for pattern in allow_patterns:
+        hf_weights_files += glob.glob(os.path.join(hf_folder, pattern))
+    use_safetensors = False
+
+    return hf_folder, hf_weights_files, use_safetensors
+
+
+EntryClass = Grok1ModelForCausalLM
\ No newline at end of file

python/sglang/srt/utils.py

@@ -11,13 +11,13 @@ from io import BytesIO
 from typing import List, Optional
 
 import numpy as np
-import pydantic
 import requests
 import torch
+import triton
 from fastapi.responses import JSONResponse
 from packaging import version as pkg_version
-from pydantic import BaseModel
 from starlette.middleware.base import BaseHTTPMiddleware
+import torch.distributed as dist
 
 logger = logging.getLogger(__name__)
 
@@ -178,7 +178,8 @@ def get_int_token_logit_bias(tokenizer, vocab_size):
 
 def wrap_kernel_launcher(kernel):
     """A faster launcher for triton kernels."""
-    import torch.distributed as dist
+    if int(triton.__version__.split(".")[0]) >= 3:
+        return None
 
     if dist.is_initialized():
         rank = dist.get_rank()

python/sglang/test/test_utils.py

@@ -95,6 +95,11 @@ def call_generate_xinfer(prompt, temperature, max_tokens, stop=None, url=None):
     sampler_channel = grpc.insecure_channel(url.replace("http://", ""))
     sampler = sampler_pb2_grpc.SamplerStub(sampler_channel)
 
+    if stop is None:
+        stop_strings = None
+    else:
+        stop_strings = [stop]
+
     sample_request = sampler_pb2.SampleTextRequest(
         prompt=prompt,
         settings=sampler_pb2.SampleSettings(
@@ -102,7 +107,7 @@ def call_generate_xinfer(prompt, temperature, max_tokens, stop=None, url=None):
             rng_seed=0,
             temperature=max(temperature, 1e-7),
             nucleus_p=1,
-            stop_strings=[stop],
+            stop_strings=stop_strings,
         ),
     )
     stream = sampler.SampleText(sample_request)