| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import torch |
| 5 | |
| 6 | from vllm.platforms import current_platform |
| 7 | |
| 8 | |
| 9 | def _merge_attn_states_torch( |
| 10 | output: torch.Tensor, |
| 11 | prefix_output: torch.Tensor, |
| 12 | prefix_lse: torch.Tensor, |
| 13 | suffix_output: torch.Tensor, |
| 14 | suffix_lse: torch.Tensor, |
| 15 | output_lse: torch.Tensor | None = None, |
| 16 | prefill_tokens_with_context: int | None = None, |
| 17 | output_scale: torch.Tensor | None = None, |
| 18 | ) -> None: |
| 19 | assert output_scale is None, "CPU fallback does not support FP8 output" |
| 20 | |
| 21 | num_tokens = output.shape[0] |
| 22 | if prefill_tokens_with_context is None: |
| 23 | prefill_tokens_with_context = num_tokens |
| 24 | |
| 25 | prefix_lse = torch.where( |
| 26 | torch.isinf(prefix_lse) & (prefix_lse > 0), |
| 27 | torch.full_like(prefix_lse, float("-inf")), |
| 28 | prefix_lse, |
| 29 | ) |
| 30 | suffix_lse = torch.where( |
| 31 | torch.isinf(suffix_lse) & (suffix_lse > 0), |
| 32 | torch.full_like(suffix_lse, float("-inf")), |
| 33 | suffix_lse, |
| 34 | ) |
| 35 | |
| 36 | if prefill_tokens_with_context < num_tokens: |
| 37 | output[prefill_tokens_with_context:].copy_( |
| 38 | suffix_output[prefill_tokens_with_context:] |
| 39 | ) |
| 40 | if output_lse is not None: |
| 41 | output_lse[:, prefill_tokens_with_context:].copy_( |
| 42 | suffix_lse[:, prefill_tokens_with_context:] |
| 43 | ) |
| 44 | |
| 45 | if prefill_tokens_with_context == 0: |
| 46 | return |
| 47 | |
| 48 | prefix_lse = prefix_lse[:, :prefill_tokens_with_context].float() |
| 49 | suffix_lse = suffix_lse[:, :prefill_tokens_with_context].float() |
| 50 | max_lse = torch.maximum(prefix_lse, suffix_lse) |
| 51 | prefix_se = torch.exp(prefix_lse - max_lse) |
| 52 | suffix_se = torch.exp(suffix_lse - max_lse) |
| 53 | out_se = prefix_se + suffix_se |
| 54 | both_empty = max_lse == float("-inf") |
| 55 | |
| 56 | prefix_scale = (prefix_se / out_se).transpose(0, 1).unsqueeze(-1) |
| 57 | suffix_scale = (suffix_se / out_se).transpose(0, 1).unsqueeze(-1) |
| 58 | merged = ( |
| 59 | prefix_output[:prefill_tokens_with_context].float() * prefix_scale |
| 60 | + suffix_output[:prefill_tokens_with_context].float() * suffix_scale |
| 61 | ) |
| 62 | merged.masked_fill_(both_empty.transpose(0, 1).unsqueeze(-1), 0.0) |
| 63 | output[:prefill_tokens_with_context].copy_(merged.to(output.dtype)) |
| 64 | |
| 65 | if output_lse is not None: |
| 66 | merged_lse = torch.log(out_se) + max_lse |
| 67 | merged_lse.masked_fill_(both_empty, float("-inf")) |
| 68 | output_lse[:, :prefill_tokens_with_context].copy_(merged_lse) |
| 69 | |
| 70 | |
| 71 | def merge_attn_states( |
| 72 | output: torch.Tensor, |
| 73 | prefix_output: torch.Tensor, |
| 74 | prefix_lse: torch.Tensor, |
| 75 | suffix_output: torch.Tensor, |
| 76 | suffix_lse: torch.Tensor, |
| 77 | output_lse: torch.Tensor | None = None, |
| 78 | prefill_tokens_with_context: int | None = None, |
| 79 | output_scale: torch.Tensor | None = None, |
| 80 | ) -> None: |
| 81 | """Merge partial attention outputs from prefix (KV cache) and suffix |
| 82 | (new tokens) into a single output tensor using the log-sum-exp (LSE) |
| 83 | rescaling method described in section 2.2 of |
| 84 | https://www.arxiv.org/pdf/2501.01005. |
| 85 | |
| 86 | For tokens that have prefix context (token index < prefill_tokens_with_context), |
| 87 | the prefix and suffix partial outputs are combined as a weighted sum. |
| 88 | For tokens without prefix context, the suffix output is copied directly. |
| 89 | |
| 90 | Args: |
| 91 | output: Output tensor of shape [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]. |
| 92 | prefix_output: Partial attention output over the prefix (KV cache), |
| 93 | shape [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]. |
| 94 | prefix_lse: Log-sum-exp values for the prefix attention, |
| 95 | shape [NUM_HEADS, NUM_TOKENS]. |
| 96 | suffix_output: Partial attention output over the suffix (new KV), |
| 97 | shape [NUM_TOKENS, NUM_HEADS, HEAD_SIZE]. |
| 98 | suffix_lse: Log-sum-exp values for the suffix attention, |
| 99 | shape [NUM_HEADS, NUM_TOKENS]. |
| 100 | output_lse: Optional tensor to store the merged LSE values, |
| 101 | shape [NUM_HEADS, NUM_TOKENS]. If None, LSE is not written out. |
| 102 | prefill_tokens_with_context: Number of prefill tokens that have |
| 103 | prefix context and therefore require merging. Tokens at indices |
| 104 | >= this value are decode or context-free prefill tokens whose |
| 105 | output is taken directly from suffix_output. If None, all tokens |
| 106 | are treated as having context. |
| 107 | output_scale: Optional scalar tensor for FP8 static quantization. |
| 108 | When provided, output must be FP8 dtype. |
| 109 | |
| 110 | """ |
| 111 | # Both the CUDA and Triton kernels derive the suffix head stride from |
| 112 | # prefix_output, so suffix_output must share the same head stride. |
| 113 | assert prefix_output.stride(1) == suffix_output.stride(1), ( |
| 114 | "merge_attn_states requires prefix_output and suffix_output to have " |
| 115 | f"matching head strides, got {prefix_output.stride(1)} and " |
| 116 | f"{suffix_output.stride(1)}" |
| 117 | ) |
| 118 | |
| 119 | # NOTE(DefTruth): Currently, custom merge_attn_states CUDA kernel |
| 120 | # does not support FP8 dtype for inputs, fallback to use Triton kernel. |
| 121 | # However, when output_scale is provided, the inputs are still BF16/FP16 |
| 122 | # and the output is FP8 — both CUDA and Triton support this. |
| 123 | # FP8 output requires output_scale to be set. |
| 124 | if output.dtype not in (torch.float32, torch.half, torch.bfloat16): |
| 125 | assert output_scale is not None, ( |
| 126 | f"output_scale is required when output is {output.dtype}" |
| 127 | ) |
| 128 | |
| 129 | def supported_dtypes(prefix: torch.Tensor) -> bool: |
| 130 | return prefix.dtype in [torch.float32, torch.half, torch.bfloat16] |
| 131 | |
| 132 | # NOTE(DefTruth): Currently, custom merge_attn_states CUDA |
| 133 | # kernel load/store 128b(16 bytes) per memory issue within |
| 134 | # thread. Namely, the headsize(headdim) must be multiple of |
| 135 | # pack_size based on input dtype (float32 -> 4, half/bfloat16 -> 8). |
| 136 | def supported_headdim(prefix: torch.Tensor) -> bool: |
| 137 | headdim = prefix.shape[2] # [NUM_TOKENS, NUM_HEADS, HEAD_SIZE] |
| 138 | if prefix.dtype == torch.float32: |
| 139 | return headdim % 4 == 0 |
| 140 | return headdim % 8 == 0 |
| 141 | |
| 142 | if ( |
| 143 | current_platform.is_cuda() |
| 144 | and supported_dtypes(prefix_output) |
| 145 | and supported_headdim(prefix_output) |
| 146 | ): |
| 147 | from vllm._custom_ops import merge_attn_states |
| 148 | |
| 149 | return merge_attn_states( |
| 150 | output, |
| 151 | prefix_output, |
| 152 | prefix_lse, |
| 153 | suffix_output, |
| 154 | suffix_lse, |
| 155 | output_lse, |
| 156 | prefill_tokens_with_context, |
| 157 | output_scale, |
| 158 | ) |
| 159 | elif current_platform.is_cpu(): |
| 160 | return _merge_attn_states_torch( |
| 161 | output, |
| 162 | prefix_output, |
| 163 | prefix_lse, |
| 164 | suffix_output, |
| 165 | suffix_lse, |
| 166 | output_lse, |
| 167 | prefill_tokens_with_context, |
| 168 | output_scale, |
| 169 | ) |
| 170 | else: |
| 171 | from vllm.v1.attention.ops.triton_merge_attn_states import ( |
| 172 | merge_attn_states, |
| 173 | ) |
| 174 | |
| 175 | return merge_attn_states( |
| 176 | output, |
| 177 | prefix_output, |
| 178 | prefix_lse, |
| 179 | suffix_output, |
| 180 | suffix_lse, |
| 181 | output_lse, |
| 182 | prefill_tokens_with_context, |
| 183 | output_scale, |
| 184 | ) |