| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Inkling logits processor (muP + LoRA aware). |
| 4 | |
| 5 | Inkling divides the final logits by a muP width multiplier |
| 6 | (``logits_mup_width_multiplier``). This applies it two ways, depending on |
| 7 | whether an lm_head LoRA is attached: |
| 8 | |
| 9 | * No LoRA: fold ``1/mup`` into the lm_head GEMM alpha (fp32 epilogue) -- no |
| 10 | separate elementwise kernel, no extra rounding, no weight mutation. |
| 11 | * LoRA attached: the LoRA manager wraps this layer in |
| 12 | ``LogitsProcessorWithLoRA``, whose ``forward`` calls |
| 13 | ``type(base_layer).forward(self=wrapper)`` -- so this ``forward`` runs with |
| 14 | ``self`` bound to the wrapper. We detect that via ``base_layer`` and take the |
| 15 | LoRA path: run the wrapper's ``_get_logits`` (base logits + the lm_head LoRA |
| 16 | delta), then divide the full logits by the multiplier so the delta is scaled |
| 17 | too. muP thus composes with the LoRA delta, with the dispatch as the only |
| 18 | model-side branching. |
| 19 | """ |
| 20 | |
| 21 | from __future__ import annotations |
| 22 | |
| 23 | import torch |
| 24 | |
| 25 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 26 | from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding |
| 27 | |
| 28 | |
| 29 | class InklingLogitsProcessor(LogitsProcessor): |
| 30 | """``LogitsProcessor`` that applies Inkling's muP logits width multiplier. |
| 31 | |
| 32 | Args: |
| 33 | vocab_size: Padded vocabulary size. |
| 34 | org_vocab_size: Unpadded vocabulary size (defaults to ``vocab_size``). |
| 35 | scale: Base logits scale (kept ``1.0`` for the served checkpoint). |
| 36 | logits_as_input: Whether the input is already logits. |
| 37 | soft_cap: Optional logit soft cap (``None`` for the served checkpoint). |
| 38 | logits_mup_width_multiplier: muP width divisor for the final logits; |
| 39 | ``None`` or ``0`` disables it. |
| 40 | |
| 41 | """ |
| 42 | |
| 43 | def __init__( |
| 44 | self, |
| 45 | vocab_size: int, |
| 46 | org_vocab_size: int | None = None, |
| 47 | scale: float = 1.0, |
| 48 | logits_as_input: bool = False, |
| 49 | soft_cap: float | None = None, |
| 50 | logits_mup_width_multiplier: float | None = None, |
| 51 | ) -> None: |
| 52 | super().__init__( |
| 53 | vocab_size=vocab_size, |
| 54 | org_vocab_size=org_vocab_size, |
| 55 | scale=scale, |
| 56 | logits_as_input=logits_as_input, |
| 57 | soft_cap=soft_cap, |
| 58 | ) |
| 59 | self.logits_mup_width_multiplier = logits_mup_width_multiplier |
| 60 | self._logits_zero: torch.Tensor | None = None |
| 61 | |
| 62 | def forward( |
| 63 | self, |
| 64 | lm_head: VocabParallelEmbedding, |
| 65 | hidden_states: torch.Tensor, |
| 66 | embedding_bias: torch.Tensor | None = None, |
| 67 | skip_gather: bool = False, |
| 68 | ) -> torch.Tensor | None: |
| 69 | # ``base_layer`` exists only on the LogitsProcessorWithLoRA wrapper, |
| 70 | # which calls this forward with ``self`` bound to the wrapper. The |
| 71 | # wrapper is not an ``InklingLogitsProcessor`` instance, so dispatch |
| 72 | # ``_lora_forward`` explicitly through the base_layer's class (it |
| 73 | # provides ``_get_logits``/``logits_as_input``; only ``_lora_forward`` |
| 74 | # lives on this class). |
| 75 | if hasattr(self, "base_layer"): |
| 76 | return type(self.base_layer)._lora_forward( |
| 77 | self, lm_head, hidden_states, embedding_bias, skip_gather |
| 78 | ) |
| 79 | return self._base_forward(lm_head, hidden_states, embedding_bias, skip_gather) |
| 80 | |
| 81 | def _lora_forward( |
| 82 | self, |
| 83 | lm_head: VocabParallelEmbedding, |
| 84 | hidden_states: torch.Tensor, |
| 85 | embedding_bias: torch.Tensor | None = None, |
| 86 | skip_gather: bool = False, |
| 87 | ) -> torch.Tensor | None: |
| 88 | # ``self`` is the LogitsProcessorWithLoRA wrapper here: ``_get_logits`` |
| 89 | # returns the base logits plus the lm_head LoRA delta. Apply the muP |
| 90 | # divisor on the full logits so the LoRA delta is scaled too. |
| 91 | mup_multiplier = self.base_layer.logits_mup_width_multiplier |
| 92 | mup = 1.0 / mup_multiplier if mup_multiplier else None |
| 93 | if self.logits_as_input: |
| 94 | logits = hidden_states |
| 95 | else: |
| 96 | logits = self._get_logits( |
| 97 | hidden_states, lm_head, embedding_bias, skip_gather |
| 98 | ) |
| 99 | # TODO: fuse this multiplication |
| 100 | if logits is not None and mup: |
| 101 | assert self.base_layer.soft_cap is None |
| 102 | assert self.base_layer.scale == 1.0 |
| 103 | logits = logits * mup |
| 104 | return logits |
| 105 | |
| 106 | def _base_forward( |
| 107 | self, |
| 108 | lm_head: VocabParallelEmbedding, |
| 109 | hidden_states: torch.Tensor, |
| 110 | embedding_bias: torch.Tensor | None = None, |
| 111 | skip_gather: bool = False, |
| 112 | ) -> torch.Tensor | None: |
| 113 | mup = self.logits_mup_width_multiplier |
| 114 | if not mup: |
| 115 | return super().forward(lm_head, hidden_states, embedding_bias, skip_gather) |
| 116 | # Fold the muP width divisor into the lm_head GEMM alpha (fp32 epilogue): |
| 117 | # no separate elementwise kernel, no bf16 rounding of scaled logits, and |
| 118 | # no weight mutation. Overfit to the served checkpoint: bf16 lm_head, no |
| 119 | # soft cap, unit logits scale. |
| 120 | assert self.soft_cap is None |
| 121 | assert self.scale == 1.0 |
| 122 | w = lm_head.weight |
| 123 | if self._logits_zero is None: |
| 124 | self._logits_zero = w.new_zeros(1) |
| 125 | logits = torch.addmm( |
| 126 | self._logits_zero, |
| 127 | hidden_states, |
| 128 | w.t(), |
| 129 | beta=0.0, |
| 130 | alpha=1.0 / mup, |
| 131 | ) |
| 132 | if skip_gather: |
| 133 | return logits |
| 134 | logits = self._gather_logits(logits) |
| 135 | if logits is not None: |
| 136 | logits = logits[..., : self.org_vocab_size] |
| 137 | return logits |