已合并
[pytorch][bugfix] align minicpm loss #3297
wx_4e191bb7af创建于 2025年9月11日
[pytorch][bugfix] align minicpm loss #3297
已合并
从refs/pull/3297/head合入到master
共 3 个文件变更+132-2
| @@ -15,4 +15,5 @@ python convert_ckpt.py \ | |||
| 15 | --target-pipeline-parallel-size 4 \ | 15 | --target-pipeline-parallel-size 4 \ |
| 16 | --target-expert-parallel-size 2 \ | 16 | --target-expert-parallel-size 2 \ |
| 17 | --model-type-hf minicpm-moe \ | 17 | --model-type-hf minicpm-moe \ |
| 18 | + --moe-grouped-gemm \ | ||
| 18 | --params-dtype bf16 | 19 | --params-dtype bf16 |
| @@ -24,8 +24,10 @@ MOE_ARGS=" | |||
| 24 | --moe-router-topk 2 \ | 24 | --moe-router-topk 2 \ |
| 25 | --moe-router-load-balancing-type aux_loss \ | 25 | --moe-router-load-balancing-type aux_loss \ |
| 26 | --moe-aux-loss-coeff 0.01 \ | 26 | --moe-aux-loss-coeff 0.01 \ |
| 27 | - --moe-token-dispatcher-type allgather \ | 27 | + --moe-token-dispatcher-type alltoall_seq \ |
| 28 | + --moe-alltoall-overlap-comm \ | ||
| 28 | --moe-permutation-async-comm \ | 29 | --moe-permutation-async-comm \ |
| 30 | + --moe-permute-fusion \ | ||
| 29 | --moe-grouped-gemm \ | 31 | --moe-grouped-gemm \ |
| 30 | --moe-layer-freq -1 \ | 32 | --moe-layer-freq -1 \ |
| 31 | --first-k-dense-replace -1 \ | 33 | --first-k-dense-replace -1 \ |
| @@ -56,6 +58,7 @@ GPT_ARGS=" | |||
| 56 | --micro-batch-size 1 \ | 58 | --micro-batch-size 1 \ |
| 57 | --global-batch-size 128 \ | 59 | --global-batch-size 128 \ |
| 58 | --make-vocab-size-divisible-by 1 \ | 60 | --make-vocab-size-divisible-by 1 \ |
| 61 | + --gemm-gradient-accumulation-fusion \ | ||
| 59 | --lr 1.25e-5 \ | 62 | --lr 1.25e-5 \ |
| 60 | --train-iters 5000 \ | 63 | --train-iters 5000 \ |
| 61 | --lr-decay-style cosine \ | 64 | --lr-decay-style cosine \ |
| @@ -78,7 +81,6 @@ GPT_ARGS=" | |||
| 78 | --adam-beta1 0.9 \ | 81 | --adam-beta1 0.9 \ |
| 79 | --initial-loss-scale 65536 \ | 82 | --initial-loss-scale 65536 \ |
| 80 | --adam-beta2 0.95 \ | 83 | --adam-beta2 0.95 \ |
| 81 | - --no-gradient-accumulation-fusion \ | ||
| 82 | --no-load-optim \ | 84 | --no-load-optim \ |
| 83 | --no-load-rng \ | 85 | --no-load-rng \ |
| 84 | --use-distributed-optimizer \ | 86 | --use-distributed-optimizer \ |
| @@ -14,12 +14,15 @@ | |||
| 14 | # limitations under the License. | 14 | # limitations under the License. |
| 15 | 15 | ||
| 16 | import math | 16 | import math |
| 17 | +from typing import Any, Dict, Optional, Tuple | ||
| 18 | +from torch import Tensor | ||
| 17 | 19 | ||
| 18 | from megatron.core import tensor_parallel | 20 | from megatron.core import tensor_parallel |
| 19 | from megatron.core.transformer.transformer_layer import TransformerLayerSubmodules | 21 | from megatron.core.transformer.transformer_layer import TransformerLayerSubmodules |
| 20 | from megatron.core.utils import WrappedTensor, deprecate_inference_params | 22 | from megatron.core.utils import WrappedTensor, deprecate_inference_params |
| 21 | from megatron.core.transformer.transformer_layer import TransformerLayer as MegatronTransformerLayer | 23 | from megatron.core.transformer.transformer_layer import TransformerLayer as MegatronTransformerLayer |
| 22 | from megatron.core.transformer.transformer_config import TransformerConfig | 24 | from megatron.core.transformer.transformer_config import TransformerConfig |
| 25 | +from megatron.core.packed_seq_params import PackedSeqParams | ||
| 23 | from megatron.core.transformer.moe.moe_layer import MoELayer | 26 | from megatron.core.transformer.moe.moe_layer import MoELayer |
| 24 | from megatron.core.transformer.moe.experts import GroupedMLP, SequentialMLP | 27 | from megatron.core.transformer.moe.experts import GroupedMLP, SequentialMLP |
| 25 | from megatron.core.utils import make_viewless_tensor | 28 | from megatron.core.utils import make_viewless_tensor |
| @@ -58,6 +61,130 @@ class TransformerLayer(MegatronTransformerLayer): | |||
| 58 | self.mtp_idx = 0 | 61 | self.mtp_idx = 0 |
| 59 | self.self_attention.core_attention.mtp_idx = 0 | 62 | self.self_attention.core_attention.mtp_idx = 0 |
| 60 | 63 | ||
| 64 | + def _forward_attention( | ||
| 65 | + self, | ||
| 66 | + hidden_states: Tensor, | ||
| 67 | + attention_mask: Optional[Tensor] = None, | ||
| 68 | + context: Optional[Tensor] = None, | ||
| 69 | + context_mask: Optional[Tensor] = None, | ||
| 70 | + rotary_pos_emb: Optional[Tensor] = None, | ||
| 71 | + rotary_pos_cos: Optional[Tensor] = None, | ||
| 72 | + rotary_pos_sin: Optional[Tensor] = None, | ||
| 73 | + attention_bias: Optional[Tensor] = None, | ||
| 74 | + inference_context: Optional[Any] = None, | ||
| 75 | + packed_seq_params: Optional[PackedSeqParams] = None, | ||
| 76 | + sequence_len_offset: Optional[Tensor] = None, | ||
| 77 | + *, | ||
| 78 | + inference_params: Optional[Any] = None, | ||
| 79 | + ): | ||
| 80 | + """ | ||
| 81 | + Perform a forward pass through the attention layer and the layernorms before and after | ||
| 82 | + the attention operations. | ||
| 83 | + | ||
| 84 | + Args: | ||
| 85 | + hidden_states (Tensor): Input tensor of shape [s, b, h] where s is sequence length, | ||
| 86 | + b is batch size, and h is hidden size. | ||
| 87 | + attention_mask (Tensor): Mask tensor for self-attention. | ||
| 88 | + context (Tensor, optional): Context tensor for cross-attention. | ||
| 89 | + context_mask (Tensor, optional): Mask tensor for cross-attention. | ||
| 90 | + rotary_pos_emb (Tensor, optional): Rotary positional embeddings. | ||
| 91 | + attention_bias (Tensor, optional): Bias tensor for Q * K.T. | ||
| 92 | + inference_context (object, optional): Parameters for inference-time optimizations. | ||
| 93 | + packed_seq_params (object, optional): Parameters for packed sequence processing. | ||
| 94 | + sequence_len_offset (Tensor, optional): Offset along sequence dimension | ||
| 95 | + during inference. | ||
| 96 | + | ||
| 97 | + Returns: | ||
| 98 | + Tuple[Tensor, Tensor, Tensor]: A tuple containing: | ||
| 99 | + pre_mlp_layernorm_output (Tensor): Transformed hidden states before the MLP. | ||
| 100 | + residual (Tensor): Residual connection. | ||
| 101 | + context (Tensor): Updated context tensor if cross-attention is used, | ||
| 102 | + otherwise None. | ||
| 103 | + """ | ||
| 104 | + args = get_args() | ||
| 105 | + inference_context = deprecate_inference_params(inference_context, inference_params) | ||
| 106 | + | ||
| 107 | + # Residual connection. | ||
| 108 | + residual = hidden_states | ||
| 109 | + | ||
| 110 | + # Optional Input Layer norm | ||
| 111 | + if self.recompute_input_layernorm: | ||
| 112 | + self.input_layernorm_checkpoint = tensor_parallel.CheckpointWithoutOutput() | ||
| 113 | + input_layernorm_output = self.input_layernorm_checkpoint.checkpoint( | ||
| 114 | + self.input_layernorm, hidden_states | ||
| 115 | + ) | ||
| 116 | + else: | ||
| 117 | + input_layernorm_output = self.input_layernorm(hidden_states) | ||
| 118 | + | ||
| 119 | + # Self attention. | ||
| 120 | + attention_output_with_bias = self.self_attention( | ||
| 121 | + input_layernorm_output, | ||
| 122 | + attention_mask=attention_mask, | ||
| 123 | + inference_context=inference_context, | ||
| 124 | + rotary_pos_emb=rotary_pos_emb, | ||
| 125 | + rotary_pos_cos=rotary_pos_cos, | ||
| 126 | + rotary_pos_sin=rotary_pos_sin, | ||
| 127 | + attention_bias=attention_bias, | ||
| 128 | + packed_seq_params=packed_seq_params, | ||
| 129 | + sequence_len_offset=sequence_len_offset, | ||
| 130 | + ) | ||
| 131 | + | ||
| 132 | + # For minicpm model | ||
| 133 | + if args.scale_depth is not None: | ||
| 134 | + attention_output, attention_bias = attention_output_with_bias | ||
| 135 | + attention_output = attention_output * (args.scale_depth / math.sqrt(args.num_layers)) | ||
| 136 | + attention_output_with_bias = (attention_output, attention_bias) | ||
| 137 | + | ||
| 138 | + if self.recompute_input_layernorm: | ||
| 139 | + # discard the output of the input layernorm and register the recompute | ||
| 140 | + # as a gradient hook of attention_output_with_bias[0] | ||
| 141 | + self.input_layernorm_checkpoint.discard_output_and_register_recompute( | ||
| 142 | + attention_output_with_bias[0] | ||
| 143 | + ) | ||
| 144 | + | ||
| 145 | + # inside the module provided in the `bias_dropout_add_spec` module? | ||
| 146 | + with self.bias_dropout_add_exec_handler(): | ||
| 147 | + hidden_states = self.self_attn_bda(self.training, self.config.bias_dropout_fusion)( | ||
| 148 | + attention_output_with_bias, residual, self.hidden_dropout | ||
| 149 | + ) | ||
| 150 | + | ||
| 151 | + # Residual connection. | ||
| 152 | + residual = hidden_states | ||
| 153 | + | ||
| 154 | + # Optional Layer norm after self-attention | ||
| 155 | + pre_cross_attn_layernorm_output = self.pre_cross_attn_layernorm(hidden_states) | ||
| 156 | + | ||
| 157 | + # Cross attention. | ||
| 158 | + attention_output_with_bias = self.cross_attention( | ||
| 159 | + pre_cross_attn_layernorm_output, | ||
| 160 | + attention_mask=context_mask, | ||
| 161 | + key_value_states=context, | ||
| 162 | + inference_context=inference_context, | ||
| 163 | + ) | ||
| 164 | + | ||
| 165 | + if isinstance(attention_output_with_bias, dict) and "context" in attention_output_with_bias: | ||
| 166 | + context = attention_output_with_bias["context"] | ||
| 167 | + | ||
| 168 | + # inside the module provided in the `bias_dropout_add_spec` module? | ||
| 169 | + with self.bias_dropout_add_exec_handler(): | ||
| 170 | + hidden_states = self.cross_attn_bda(self.training, self.config.bias_dropout_fusion)( | ||
J | |||
| 171 | + attention_output_with_bias, residual, self.hidden_dropout | ||
W
megatron原生逻辑   | |||
| 172 | + ) | ||
| 173 | + | ||
| 174 | + # Residual connection. | ||
| 175 | + residual = hidden_states | ||
| 176 | + | ||
| 177 | + # Optional Layer norm post the cross-attention. | ||
| 178 | + if self.recompute_pre_mlp_layernorm: | ||
| 179 | + self.pre_mlp_norm_checkpoint = tensor_parallel.CheckpointWithoutOutput() | ||
| 180 | + pre_mlp_layernorm_output = self.pre_mlp_norm_checkpoint.checkpoint( | ||
W
适配minicpm模型scale_depth特性 | |||
| 181 | + self.pre_mlp_layernorm, hidden_states | ||
| 182 | + ) | ||
| 183 | + else: | ||
| 184 | + pre_mlp_layernorm_output = self.pre_mlp_layernorm(hidden_states) | ||
| 185 | + | ||
| 186 | + return pre_mlp_layernorm_output, residual, context | ||
| 187 | + | ||
| 61 | def _forward_mlp(self, pre_mlp_layernorm_output, residual): | 188 | def _forward_mlp(self, pre_mlp_layernorm_output, residual): |
| 62 | args = get_args() | 189 | args = get_args() |
| 63 | # MLP. | 190 | # MLP. |
该代码无调用，无引用，用途是什么？