From c0233fcf0100c07820d7f7a725de458a0e14df64 Mon Sep 17 00:00:00 2001 From: Woosuk Kwon Date: Wed, 19 Aug 2026 19:39:52 -0700 Subject: [PATCH] [Model] Remove unused DeepseekV32Indexer forward (#53021) Signed-off-by: Woosuk Kwon Co-authored-by: OpenAI Codex --- vllm/models/deepseek_v32/attention.py | 51 --------------------------- 1 file changed, 51 deletions(-) diff --git a/vllm/models/deepseek_v32/attention.py b/vllm/models/deepseek_v32/attention.py index 4edcae4640..87e7ec24e3 100644 --- a/vllm/models/deepseek_v32/attention.py +++ b/vllm/models/deepseek_v32/attention.py @@ -21,9 +21,6 @@ from vllm.model_executor.layers.linear import ( RowParallelLinear, ) from vllm.model_executor.layers.quantization import QuantizationConfig -from vllm.model_executor.layers.quantization.utils.fp8_utils import ( - per_token_group_quant_fp8, -) from vllm.model_executor.layers.rotary_embedding import get_rope from vllm.model_executor.layers.sparse_attn_indexer import ( SparseAttnIndexer, @@ -117,54 +114,6 @@ class DeepseekV32Indexer(nn.Module): self.topk_indices_buffer, ) - def forward( - self, - hidden_states: torch.Tensor, - qr: torch.Tensor, - positions: torch.Tensor, - rotary_emb: nn.Module, - ) -> torch.Tensor: - q, _ = self.wq_b(qr) - q = q.view(-1, self.n_head, self.head_dim) - - q_pe, q_nope = torch.split( - q, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 - ) - - kw, _ = self.wk_weights_proj(hidden_states) - k = kw[:, : self.head_dim] - weights = kw[:, self.head_dim :] - - k = self.k_norm(k) - k_pe, k_nope = torch.split( - k, [self.rope_dim, self.head_dim - self.rope_dim], dim=-1 - ) - - q_pe, k_pe = rotary_emb(positions, q_pe, k_pe.unsqueeze(1)) - - q_pe = q_pe.reshape(-1, self.n_head, self.rope_dim) - k_pe = k_pe.reshape(-1, 1, self.rope_dim) - - q = torch.cat([q_pe, q_nope], dim=-1) - k = torch.cat([k_pe.squeeze(-2), k_nope], dim=-1) - - q = q.view(-1, self.head_dim) - q_fp8, q_scale = per_token_group_quant_fp8( - q, - self.quant_block_size, - column_major_scales=False, - use_ue8m0=self.scale_fmt is not None, - ) - q_fp8 = q_fp8.view(-1, self.n_head, self.head_dim) - q_scale = q_scale.view(-1, self.n_head, 1) - - weights = ( - weights.unsqueeze(-1) * q_scale * self.softmax_scale * self.n_head**-0.5 - ) - weights = weights.squeeze(-1) - - return self.indexer_op(hidden_states, q_fp8, k, weights) - class DeepseekV32Attention(MLAAttention): indexer: "DeepseekV32Indexer | None"