diff --git a/vllm/model_executor/layers/fused_moe/router/fused_topk_bias_router.py b/vllm/model_executor/layers/fused_moe/router/fused_topk_bias_router.py index bc019a58e2..f1c2636156 100644 --- a/vllm/model_executor/layers/fused_moe/router/fused_topk_bias_router.py +++ b/vllm/model_executor/layers/fused_moe/router/fused_topk_bias_router.py @@ -3,7 +3,6 @@ import functools import torch -import torch.nn.functional as F import vllm._custom_ops as ops import vllm.envs as envs @@ -72,52 +71,6 @@ def vllm_topk_sigmoid( return topk_weights, topk_indices -def _topk_softplus_sqrt_torch( - topk_weights: torch.Tensor, - topk_indices: torch.Tensor, - token_expert_indices: torch.Tensor, - gating_output: torch.Tensor, - renormalize: bool = False, - e_score_correction_bias: torch.Tensor | None = None, - input_tokens: torch.Tensor | None = None, - hash_indices_table: torch.Tensor | None = None, - routed_scaling_factor: float = 1.0, -) -> tuple[torch.Tensor, ...]: - """Pure PyTorch fallback for topk_softplus_sqrt (XPU/CPU).""" - # scores = sqrt(softplus(gating_output)) - scores = torch.sqrt(F.softplus(gating_output.float())) - - # Bias is used for expert SELECTION only, not for weight computation. - # Using biased scores as weights flattens the distribution when the bias - # is near-uniform (e.g., DSv4-Flash where all biases ≈ 8.08). - if e_score_correction_bias is not None: - scores_for_choice = scores + e_score_correction_bias.float() - else: - scores_for_choice = scores - - topk = topk_weights.shape[-1] - - if hash_indices_table is not None and input_tokens is not None: - # Hash MoE: expert indices predetermined by lookup table - # hash_indices_table: [vocab_size, topk] mapping token_id -> expert_ids - expert_ids = hash_indices_table[input_tokens.long()] # [M, topk] - topk_indices.copy_(expert_ids) - # Gather weights from unbiased scores - weights = scores.gather(1, expert_ids.long()) - else: - # Standard topk selection using biased scores - _, indices = torch.topk(scores_for_choice, k=topk, dim=-1) - topk_indices.copy_(indices) - # Gather weights from unbiased scores - weights = scores.gather(1, indices) - - if renormalize: - weights = weights / (weights.sum(dim=-1, keepdim=True).clamp(min=1e-20)) - - topk_weights.copy_(weights * routed_scaling_factor) - return topk_weights, topk_indices - - def vllm_topk_softplus_sqrt( topk_weights: torch.Tensor, topk_indices: torch.Tensor, @@ -129,21 +82,6 @@ def vllm_topk_softplus_sqrt( hash_indices_table: torch.Tensor | None = None, routed_scaling_factor: float = 1.0, ) -> tuple[torch.Tensor, ...]: - from vllm.platforms import current_platform - - if current_platform.is_xpu(): - return _topk_softplus_sqrt_torch( - topk_weights, - topk_indices, - token_expert_indices, - gating_output, - renormalize, - e_score_correction_bias, - input_tokens, - hash_indices_table, - routed_scaling_factor, - ) - ops.topk_hash_softplus_sqrt( topk_weights, topk_indices,