# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved. """Transformer.""" from contextlib import nullcontext import math import numpy as np import torch import torch.nn.functional as F from typing import Optional from megatron import get_timers, get_args, get_retro_args, core, get_num_microbatches from .module import MegatronModule from megatron.core import parallel_state, tensor_parallel, mpu from megatron.core.enums import ModelType from megatron.model import LayerNorm, RMSNorm from megatron.model.enums import AttnMaskType, LayerType, AttnType from megatron.model.fused_softmax import FusedScaleMaskSoftmax from megatron.model.fused_bias_gelu import bias_gelu_impl from megatron.model.rotary_pos_embedding import apply_rotary_pos_emb from megatron.model.utils import attention_mask_func, openai_gelu, erf_gelu import deepspeed from deepspeed.moe.layer import MoE from deepspeed.accelerator import get_accelerator from apex.transformer.functional import ( fused_apply_rotary_pos_emb, fused_apply_rotary_pos_emb_cached, ) import pdb try: from deepspeed.sequence.layer import DistributedAttention dist_attn_supported = True except ImportError: dist_attn_supported = False try: from einops import rearrange except ImportError: rearrange = None import os #os.environ["TRITON_CACHE_DIR"] = f"/work/for_trans/cache" try: # FlashAttention (1.x) from flash_attn.flash_attn_interface import flash_attn_unpadded_func from flash_attn.flash_attn_triton import flash_attn_func except ImportError: flash_attn_unpadded_func = None flash_attn_func = None try: # FlashAttention-2 from flash_attn.flash_attn_interface import flash_attn_varlen_func except ImportError: flash_attn_varlen_func = None FlashAttentionBuilder = get_accelerator().get_op_builder("FlashAttentionBuilder") flash_attn_builder = None """ We use the following notation throughout this file: h: hidden size n: number of attention heads p: number of model parallel partitions np: n/p hp: h/p hn: h/n b: batch size s: sequence length l: number of layers Transformer takes input of size [s, b, h] and returns a tensor of the same size. We use the following arguments: hyperparameters: transformer hyperparameters """ class DropPath(MegatronModule): """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). """ def __init__(self, drop_prob=0.): super(DropPath, self).__init__() self.drop_prob = drop_prob def forward(self, hidden_state): if self.drop_prob == 0. or not self.training: return hidden_state keep_prob = 1 - self.drop_prob # work with diff dim tensors, not just 2D ConvNets # hidden_state: [s, b, h] shape = (1,) + (hidden_state.shape[1],) + (1,) * (hidden_state.ndim - 2) random_tensor = keep_prob + \ torch.rand(shape, dtype=hidden_state.dtype, device=hidden_state.device) random_tensor.floor_() # binarize output = hidden_state.div(keep_prob) * random_tensor return output class ParallelMLP(MegatronModule): """MLP. MLP will take the input with h hidden state, project it to 4*h hidden dimension, perform nonlinear transformation, and project the state back into h hidden dimension. """ def __init__(self, config, moe=False, enable_expert_tensor_parallelism=False): super(ParallelMLP, self).__init__() args = get_args() self.add_bias = config.add_bias_linear ffn_hidden_size = config.ffn_hidden_size if config.gated_linear_unit: ffn_hidden_size *= 2 # Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf self.dense_h_to_4h = tensor_parallel.ColumnParallelLinear( config.hidden_size, ffn_hidden_size, config=config, init_method=config.init_method, bias=self.add_bias, gather_output=False, skip_bias_add=True, moe=moe, enable_expert_tensor_parallelism=enable_expert_tensor_parallelism ) self.bias_gelu_fusion = False self.activation_func = None self.swiglu = args.swiglu if args.openai_gelu: self.activation_func = openai_gelu elif args.onnx_safe: self.activation_func = erf_gelu elif args.swiglu: def swiglu(x): x = torch.chunk(x, 2, dim=-1) return F.silu(x[0]) * x[1] self.activation_func = swiglu elif args.squared_relu: def squared_relu(x): return torch.pow(F.relu(x), 2) self.activation_func = squared_relu else: self.bias_gelu_fusion = args.bias_gelu_fusion self.activation_func = F.gelu # Project back to h. self.dense_4h_to_h = tensor_parallel.RowParallelLinear( config.ffn_hidden_size, config.hidden_size, config=config, init_method=config.output_layer_init_method, bias=self.add_bias, input_is_parallel=True, moe=moe, enable_expert_tensor_parallelism=enable_expert_tensor_parallelism ) def forward(self, hidden_states): # [s, b, 4hp] intermediate_parallel, bias_parallel = self.dense_h_to_4h(hidden_states) if self.bias_gelu_fusion: assert self.add_bias is True # DeepSpeed FLOPS profiler temporarily substitues functions like F.gelu to calculate the throughput assert hasattr(self, "__flops__") or self.activation_func == F.gelu intermediate_parallel = bias_gelu_impl(intermediate_parallel, bias_parallel) else: if bias_parallel is not None: intermediate_parallel = intermediate_parallel + bias_parallel intermediate_parallel = self.activation_func(intermediate_parallel) # [s, b, h] output, output_bias = self.dense_4h_to_h(intermediate_parallel) return output, output_bias class SwitchMLP(MegatronModule): """ Routes input to one of N MLP "experts" """ def __init__(self, config): super(SwitchMLP, self).__init__() args = get_args() self.router = torch.nn.Linear(config.hidden_size, args.num_experts_switch) self.experts = torch.nn.ModuleList() for i in range(args.num_experts_switch): self.experts.append(ParallelMLP(config)) def forward(self, hidden_states): # hidden_states: [s, b, h] s = hidden_states.size(0) b = hidden_states.size(1) h = hidden_states.size(2) route = self.router(hidden_states) route = torch.nn.functional.softmax(route, dim=2) max_prob, max_ind = torch.max(route, dim=2) max_prob = torch.unsqueeze(max_prob, 2) # [s b 1] # TODO (rprenger) TODO this could be made easier to read # Converting [s, b, h] to [s*b, h]. # Each vector could be routed differently hidden_states = hidden_states.view(-1, hidden_states.size(2)) # [s*b h] max_prob = max_prob.view(-1, max_prob.size(2)) # [s*b 1] max_ind = max_ind.view(-1) # [s*b] output_total = torch.empty_like(hidden_states) output_bias_total = torch.empty_like(hidden_states) #TODO (rprenger) This does each expert in serial, but it could be parallelized for expert_num, expert in enumerate(self.experts): local_indices = (max_ind == expert_num).nonzero() hidden = hidden_states[local_indices,:] output, output_bias = expert(hidden) output_bias = output_bias.expand_as(output) output_total[local_indices,:] = output output_bias_total[local_indices,:] = output_bias output_total = output_total*max_prob output_bias_total = output_bias_total*max_prob output_total = output_total.view(s, b, h) output_bias_total = output_bias_total.view(s, b, h) return output_total, output_bias_total class CoreAttention(MegatronModule): def __init__(self, layer_number, config, attn_mask_type=AttnMaskType.padding): super(CoreAttention, self).__init__() self.fp16 = config.fp16 self.bf16 = config.bf16 self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32 if self.apply_query_key_layer_scaling: self.attention_softmax_in_fp32 = True self.layer_number = max(1, layer_number) self.attn_mask_type = attn_mask_type self.sequence_parallel = config.sequence_parallel projection_size = config.kv_channels * config.num_attention_heads # Per attention head and per partition values. seq_parallel_world_size = 1 if parallel_state.sequence_parallel_is_initialized(): seq_parallel_world_size = parallel_state.get_sequence_parallel_world_size() world_size = seq_parallel_world_size if seq_parallel_world_size > 1 else parallel_state.get_tensor_model_parallel_world_size() self.hidden_size_per_partition = core.utils.divide(projection_size, world_size) self.hidden_size_per_attention_head = core.utils.divide( projection_size, config.num_attention_heads) self.num_attention_heads_per_partition = core.utils.divide( config.num_attention_heads, world_size) coeff = None self.norm_factor = math.sqrt(self.hidden_size_per_attention_head) if self.apply_query_key_layer_scaling: coeff = self.layer_number self.norm_factor *= coeff self.scale_mask_softmax = FusedScaleMaskSoftmax( self.fp16, self.bf16, self.attn_mask_type, config.masked_softmax_fusion, attention_mask_func, self.attention_softmax_in_fp32, coeff) # Dropout. Note that for a single iteration, this layer will generate # different outputs on different number of parallel partitions but # on average it should not be partition dependent. self.attention_dropout = torch.nn.Dropout(config.attention_dropout) def forward(self, query_layer, key_layer, value_layer, attention_mask): # =================================== # Raw attention scores. [b, np, s, s] # =================================== # [b, np, sq, sk] output_size = (query_layer.size(1), query_layer.size(2), query_layer.size(0), key_layer.size(0)) # [sq, b, np, hn] -> [sq, b * np, hn] query_layer = query_layer.view(output_size[2], output_size[0] * output_size[1], -1) # [sk, b, np, hn] -> [sk, b * np, hn] key_layer = key_layer.view(output_size[3], output_size[0] * output_size[1], -1) # preallocting input tensor: [b * np, sq, sk] matmul_input_buffer = parallel_state.get_global_memory_buffer().get_tensor( (output_size[0]*output_size[1], output_size[2], output_size[3]), query_layer.dtype, "mpu") # Raw attention scores. [b * np, sq, sk] matmul_result = torch.baddbmm( matmul_input_buffer, query_layer.transpose(0, 1), # [b * np, sq, hn] key_layer.transpose(0, 1).transpose(1, 2), # [b * np, hn, sk] beta=0.0, alpha=(1.0/self.norm_factor)) # change view to [b, np, sq, sk] attention_scores = matmul_result.view(*output_size) # =========================== # Attention probs and dropout # =========================== # attention scores and attention mask [b, np, sq, sk] attention_probs = self.scale_mask_softmax(attention_scores, attention_mask) # This is actually dropping out entire tokens to attend to, which might # seem a bit unusual, but is taken from the original Transformer paper. if not self.sequence_parallel: with tensor_parallel.get_cuda_rng_tracker().fork(): attention_probs = self.attention_dropout(attention_probs) else: attention_probs = self.attention_dropout(attention_probs) # ========================= # Context layer. [sq, b, hp] # ========================= # value_layer -> context layer. # [sk, b, np, hn] --> [b, np, sq, hn] # context layer shape: [b, np, sq, hn] output_size = (value_layer.size(1), value_layer.size(2), query_layer.size(0), value_layer.size(3)) # change view [sk, b * np, hn] value_layer = value_layer.view(value_layer.size(0), output_size[0] * output_size[1], -1) # change view [b * np, sq, sk] attention_probs = attention_probs.view(output_size[0] * output_size[1], output_size[2], -1) # matmul: [b * np, sq, hn] context_layer = torch.bmm(attention_probs, value_layer.transpose(0, 1)) # change view [b, np, sq, hn] context_layer = context_layer.view(*output_size) # [b, np, sq, hn] --> [sq, b, np, hn] context_layer = context_layer.permute(2, 0, 1, 3).contiguous() # [sq, b, np, hn] --> [sq, b, hp] new_context_layer_shape = context_layer.size()[:-2] + \ (self.hidden_size_per_partition,) context_layer = context_layer.view(*new_context_layer_shape) return context_layer class FlashSelfAttention(torch.nn.Module): """Implement the scaled dot product attention with softmax. Arguments --------- softmax_scale: The temperature to use for the softmax attention. (default: 1/sqrt(d_keys) where d_keys is computed at runtime) attention_dropout: The dropout rate to apply to the attention (default: 0.0) """ def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0, device=None, dtype=None): super().__init__() assert flash_attn_unpadded_func is not None or flash_attn_varlen_func is not None or flash_attn_builder is not None, \ ('Please install FlashAttention first, e.g., with pip install flash-attn or implement your own flash attention') assert rearrange is not None, 'Please install einops first, e.g., with pip install einops' self.causal = causal self.softmax_scale = softmax_scale self.dropout_p = attention_dropout # Use FlashAttention-2 when args.use_flash_attn_v2 is True args = get_args() self.flash_attn_func = flash_attn_varlen_func if args.use_flash_attn_v2 else flash_attn_unpadded_func def forward(self, q, k, v): """Implements the multihead softmax attention. Arguments --------- q, k, v: The tensor containing the query, key, and value. (B, S, H, D) """ assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q,k,v))) assert all((get_accelerator().on_accelerator(i) for i in (q, k, v))) # if get_accelerator().device_name() == 'cuda': # assert all((i.is_cuda for i in (q,k,v))) # else: # assert all((i.is_xpu for i in (q,k,v))) batch_size, seqlen_q = q.shape[0], q.shape[1] seqlen_k = k.shape[1] if get_accelerator().device_name() == 'cuda': # goes for cuda device q, k, v = [rearrange(x, 'b s ... -> (b s) ...') for x in [q, k, v]] cu_seqlens_q = torch.arange(0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32, device=q.device) else: # goes for other device q, k, v = [rearrange(x, 'b s h d -> b h s d').contiguous() for x in [q, k, v]] if self.training: # during training q,k,v always have same seqlen assert seqlen_k == seqlen_q is_causal = self.causal cu_seqlens_k = cu_seqlens_q if get_accelerator().device_name() == 'cuda' else None dropout_p = self.dropout_p else: # turn off FA causal mask after first inference autoregressive iteration # only on first autoregressive step q,k,v have same seqlen is_causal = seqlen_q == seqlen_k cu_seqlens_k = torch.arange(0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32, device=q.device) if get_accelerator().device_name() == 'cuda' else None dropout_p = 0 output = self.flash_attn_func( q, k, v, cu_seqlens_q, cu_seqlens_k, seqlen_q, seqlen_k, dropout_p, softmax_scale=self.softmax_scale, causal=is_causal ) if get_accelerator().device_name() == 'cuda' else flash_attn_builder.flash_attn_func( q, k, v, self.dropout_p, self.softmax_scale, is_causal ) output = rearrange(output, '(b s) ... -> b s ...', b=batch_size) if get_accelerator().device_name() == 'cuda' else rearrange( output, 'b h s d -> b s h d').contiguous() return output class FlashSelfAttentionTriton(torch.nn.Module): """Implement the scaled dot product attention with softmax. Arguments --------- softmax_scale: The temperature to use for the softmax attention. (default: 1/sqrt(d_keys) where d_keys is computed at runtime) attention_dropout: The dropout rate to apply to the attention (default: 0.0) """ def __init__(self, causal=False, softmax_scale=None, attention_dropout=0.0, device=None, dtype=None): super().__init__() assert flash_attn_func is not None, ('Triton version of FlashAttention is not installed.') assert rearrange is not None, 'Please install einops first, e.g., with pip install einops' self.causal = causal self.softmax_scale = softmax_scale self.dropout_p = attention_dropout def forward(self, q, k, v): """Implements the multihead softmax attention. Arguments --------- q, k, v: The tensor containing the query, key, and value. (B, S, H, D) """ assert q.dtype in [torch.float16, torch.bfloat16] assert q.is_cuda q, k, v = [rearrange(x, 's b h d -> b h s d').contiguous() for x in (q, k, v)] output = flash_attn_func(q, k, v, self.causal) output = rearrange(output, 'b s h d -> h b (s d)').contiguous() return output class ParallelAttention(MegatronModule): """Parallel self-attention layer abstract class. Self-attention layer takes input with size [s, b, h] and returns output of the same size. """ def __init__(self, config, layer_number, attention_type=AttnType.self_attn, attn_mask_type=AttnMaskType.padding): super(ParallelAttention, self).__init__() args = get_args() self.layer_number = max(1, layer_number) self.attention_type = attention_type self.attn_mask_type = attn_mask_type self.params_dtype = config.params_dtype self.sequence_parallel = config.sequence_parallel self.num_attention_heads = config.num_attention_heads self.num_key_value_heads = config.num_key_value_heads self.use_gqa = (self.num_attention_heads != self.num_key_value_heads) self.use_flash_attn = (args.use_flash_attn_v1 or args.use_flash_attn_triton or args.use_flash_attn_v2) \ and attention_type == AttnType.self_attn \ and self.attn_mask_type == AttnMaskType.causal self.use_flash_attn_triton = args.use_flash_attn_triton if self.use_flash_attn: global flash_attn_builder try: flash_attn_builder = FlashAttentionBuilder().load() except TypeError: flash_attn_builder = None if args.use_flash_attn_v1: assert flash_attn_unpadded_func != None or flash_attn_builder != None, ("Cannot import FlashAttention v1 " "and Cannot find FlashAttention Builder") if args.use_flash_attn_v2: assert flash_attn_varlen_func != None, "Cannot import FlashAttention v2 " if args.use_flash_attn_triton: assert flash_attn_func != None, "Cannot import FlashAttention triton " assert attention_type == AttnType.self_attn, ('FlashAttention code path only supports ' 'self-attention for now') assert self.attn_mask_type == AttnMaskType.causal, ('FlashAttention code path only ' 'supports causal mask for now') if rearrange is None: raise ImportError('einops is not installed, please install with pip install einops') projection_size = config.kv_channels * config.num_attention_heads # Per attention head and per partition values. world_size = parallel_state.get_tensor_model_parallel_world_size() self.hidden_size_per_attention_head = core.utils.divide( projection_size, config.num_attention_heads) self.num_attention_heads_per_partition = core.utils.divide( config.num_attention_heads, world_size) # Per GQA head and per partition values self.num_key_value_heads_per_partition = core.utils.divide( config.num_key_value_heads, world_size) self.num_key_value_groups = core.utils.divide( config.num_attention_heads, config.num_key_value_heads) kv_projection_size = config.kv_channels * config.num_key_value_heads assert self.hidden_size_per_attention_head == core.utils.divide( kv_projection_size, config.num_key_value_heads) # Strided linear layer. if attention_type == AttnType.self_attn: self.query_key_value = tensor_parallel.ColumnParallelLinear( config.hidden_size, projection_size + 2 * kv_projection_size, config=config, init_method=config.init_method, bias=args.add_bias_linear, gather_output=False) else: assert attention_type == AttnType.cross_attn self.query = tensor_parallel.ColumnParallelLinear( config.hidden_size, projection_size, config=config, init_method=config.init_method, bias=config.add_bias_linear, gather_output=False) self.key_value = tensor_parallel.ColumnParallelLinear( config.hidden_size, 2 * projection_size, config=config, init_method=config.init_method, bias=config.add_bias_linear, gather_output=False) # Currently FlashAttention only works with causal mask if self.use_flash_attn_triton: local_attn = FlashSelfAttentionTriton(causal=True, attention_dropout=args.attention_dropout) elif self.use_flash_attn: local_attn = FlashSelfAttention(causal=True, attention_dropout=config.attention_dropout) else: local_attn = CoreAttention(self.layer_number, config, self.attn_mask_type) self.enable_ds_sequence_parallel = parallel_state.get_sequence_parallel_world_size() > 1 \ or args.force_ds_sequence_parallel if self.enable_ds_sequence_parallel: assert dist_attn_supported, 'Distributed attention is not supported in this DeepSpeed version' assert args.num_attention_heads % parallel_state.get_sequence_parallel_world_size() == 0 self.dist_attn = DistributedAttention(local_attn, parallel_state.get_sequence_parallel_group()) else: if self.use_flash_attn: self.core_attention_flash = local_attn else: self.core_attention = local_attn self.checkpoint_core_attention = config.recompute_granularity == 'selective' # Output. self.dense = tensor_parallel.RowParallelLinear( projection_size, config.hidden_size, config=config, init_method=config.output_layer_init_method, bias=args.add_bias_linear, input_is_parallel=True, skip_bias_add=True) def _checkpointed_attention_forward(self, query_layer, key_layer, value_layer, attention_mask, rotary_pos_emb=None): """Forward method with activation checkpointing.""" def custom_forward(*inputs): query_layer = inputs[0] key_layer = inputs[1] value_layer = inputs[2] attention_mask = inputs[3] output_ = self.core_attention(query_layer, key_layer, value_layer, attention_mask) return output_ q_pos_emb, k_pos_emb = (None, None) if rotary_pos_emb is None \ else rotary_pos_emb hidden_states = tensor_parallel.checkpoint( custom_forward, False, query_layer, key_layer, value_layer, attention_mask, q_pos_emb, k_pos_emb) return hidden_states def _allocate_memory(self, inference_max_sequence_len, batch_size): return torch.empty( inference_max_sequence_len, batch_size, self.num_attention_heads_per_partition, self.hidden_size_per_attention_head, dtype=self.params_dtype, device=get_accelerator().current_device_name()) def repeat_kv(self, hidden_states, n_rep): slen, batch, num_key_value_heads_per_partition, head_dim = hidden_states.shape if n_rep == 1: return hidden_states hidden_states = hidden_states[:, :, :, None, :].expand( slen, batch, num_key_value_heads_per_partition, n_rep, head_dim) return hidden_states.reshape(slen, batch, num_key_value_heads_per_partition * n_rep, head_dim) def split_tensor(self, mixed_x_layer): query_layer = mixed_x_layer[:, :, :, :-2, :].reshape(mixed_x_layer.shape[:2] + (-1, self.hidden_size_per_attention_head)) key_layer = mixed_x_layer[:, :, :, -2, :] value_layer = mixed_x_layer[:, :, :, -1, :] return query_layer, key_layer, value_layer def forward(self, hidden_states, attention_mask, encoder_output=None, inference_params=None, rotary_pos_emb=None): # hidden_states: [sq, b, h] # ================================================= # Pre-allocate memory for key-values for inference. # ================================================= is_first_step = False if inference_params: if self.layer_number not in inference_params.key_value_memory_dict: inf_max_seq_len = inference_params.max_sequence_len inf_max_batch_size = inference_params.max_batch_size inference_key_memory = self._allocate_memory( inf_max_seq_len, inf_max_batch_size) inference_value_memory = self._allocate_memory( inf_max_seq_len, inf_max_batch_size) inference_params.key_value_memory_dict[self.layer_number] = ( inference_key_memory, inference_value_memory) is_first_step = True else: inference_key_memory, inference_value_memory = \ inference_params.key_value_memory_dict[self.layer_number] # ===================== # Query, Key, and Value # ===================== if self.attention_type == AttnType.self_attn: # Attention heads [sq, b, h] --> [sq, b, ((nq + 2 * nkv) * hn)] mixed_x_layer, _ = self.query_key_value(hidden_states) # [sq, b, ((nq + 2 * nkv) * hn)] --> [sq, b, nkv, (nq // nkv + 2), hn] new_tensor_shape = mixed_x_layer.size()[:-1] + \ (-1, (self.num_key_value_groups + 2), self.hidden_size_per_attention_head) mixed_x_layer = mixed_x_layer.view(*new_tensor_shape) # [sq, b, nkv, (nq // nkv + 2), hn] --> 3 [sq, b, np, hn] (query_layer, key_layer, value_layer) = self.split_tensor(mixed_x_layer) # Repeat kv if self.use_gqa: key_layer = self.repeat_kv(key_layer, self.num_key_value_groups) value_layer = self.repeat_kv(value_layer, self.num_key_value_groups) else: assert not self.use_gqa, 'GQA + cross-attn not tested yet' # Attention heads [sk, b, h] --> [sk, b, (np * 2 * hn)] mixed_kv_layer, _ = self.key_value(encoder_output) # [sk, b, (np * 2 * hn)] --> [sk, b, np, 2 * hn] new_tensor_shape = mixed_kv_layer.size()[:-1] + \ (self.num_attention_heads_per_partition, 2 * self.hidden_size_per_attention_head) mixed_kv_layer = mixed_kv_layer.view(*new_tensor_shape) # [sk, b, np, 2 * hn] --> 2 [sk, b, np, hn] (key_layer, value_layer) = tensor_parallel.split_tensor_along_last_dim(mixed_kv_layer, 2) # Attention head [sq, b, h] --> [sq, b, hp] query_layer, _ = self.query(hidden_states) # [sq, b, hp] --> [sq, b, np, hn] new_tensor_shape = query_layer.size()[:-1] + \ (self.num_attention_heads_per_partition, self.hidden_size_per_attention_head) query_layer = query_layer.view(*new_tensor_shape) # ================================== # Adjust key and value for inference # ================================== # duplicate the pos_emb for self attention if rotary_pos_emb is not None: if isinstance(rotary_pos_emb, tuple): rotary_pos_emb = rotary_pos_emb else: rotary_pos_emb = ((rotary_pos_emb,) * 2) if inference_params: batch_start = inference_params.batch_size_offset batch_end = batch_start + key_layer.size(1) assert batch_end <= inference_key_memory.size(1) sequence_start = inference_params.sequence_len_offset sequence_end = sequence_start + key_layer.size(0) assert sequence_end <= inference_key_memory.size(0) # Copy key and values. inference_key_memory[sequence_start:sequence_end, batch_start:batch_end, ...] = key_layer inference_value_memory[sequence_start:sequence_end, batch_start:batch_end, ...] = value_layer key_layer = inference_key_memory[ :sequence_end, batch_start:batch_end, ...] value_layer = inference_value_memory[ :sequence_end, batch_start:batch_end, ...] # adjust the key rotary positional embedding if rotary_pos_emb is not None: q_pos_emb, k_pos_emb = rotary_pos_emb # need to cross check this condition during inference # if not set_inference_key_value_memory: if not is_first_step: # In inference, we compute one token at a time. # Select the correct positional embedding # (only the last token in the sequence) q_pos_emb = q_pos_emb[sequence_end - 1 : sequence_end] else: # In the first forward pass of inference, # we use the entire provided prefix. # q_pos_emb here has the rope embeddings of the entire # prefix + to-be-generated output so # we slice to just the prefix. q_pos_emb = q_pos_emb[:sequence_end, :, :, :] k_pos_emb = k_pos_emb[:sequence_end, :, :, :] rotary_pos_emb = (q_pos_emb, k_pos_emb) # ================================== # core attention computation # ================================== # apply relative positional encoding (rotary embedding) if rotary_pos_emb is not None: q_pos_emb, k_pos_emb = rotary_pos_emb #defalut #query_layer = apply_rotary_pos_emb(query_layer, q_pos_emb) #key_layer = apply_rotary_pos_emb(key_layer, k_pos_emb) #use fused_apply_rotary_pos_emb_cached cos, sin = q_pos_emb.cos(), q_pos_emb.sin() query_layer = fused_apply_rotary_pos_emb_cached(query_layer, cos, sin, False) cos, sin = k_pos_emb.cos(), k_pos_emb.sin() key_layer = fused_apply_rotary_pos_emb_cached(key_layer, cos, sin, False) # TODO, can apply positional embedding to value_layer so it has # absolute positional embedding. # otherwise, only relative positional embedding takes effect # value_layer = apply_rotary_pos_emb(value_layer, k_pos_emb) if self.enable_ds_sequence_parallel: if self.use_flash_attn: if not self.use_flash_attn_triton: query_layer, key_layer, value_layer = [rearrange(x, 's b ... -> b s ...').contiguous() for x in (query_layer, key_layer, value_layer)] context_layer = self.dist_attn(query_layer, key_layer, value_layer) if not self.use_flash_attn_triton: context_layer = rearrange(context_layer, 'b s h d -> s b (h d)').contiguous() else: context_layer = self.dist_attn(query_layer, key_layer, value_layer, attention_mask) else: if self.use_flash_attn: if not self.use_flash_attn_triton: query_layer, key_layer, value_layer = [rearrange(x, 's b ... -> b s ...') for x in (query_layer, key_layer, value_layer)] if self.sequence_parallel: context_layer = self.core_attention_flash(query_layer, key_layer, value_layer) else: with tensor_parallel.get_cuda_rng_tracker().fork(): context_layer = self.core_attention_flash(query_layer, key_layer, value_layer) if not self.use_flash_attn_triton: context_layer = rearrange(context_layer, 'b s h d -> s b (h d)').contiguous() else: if self.checkpoint_core_attention: context_layer = self._checkpointed_attention_forward( query_layer, key_layer, value_layer, attention_mask) else: context_layer = self.core_attention( query_layer, key_layer, value_layer, attention_mask) # ================= # Output. [sq, b, h] # ================= output, bias = self.dense(context_layer) return output, bias def bias_dropout_add(x, bias, residual, prob, training): # type: (Tensor, Optional[Tensor], Tensor, float, bool) -> Tensor if bias is not None: x = x + bias out = torch.nn.functional.dropout(x, p=prob, training=training) out = residual + out return out def get_bias_dropout_add(training): def _bias_dropout_add(x, bias, residual, prob): return bias_dropout_add(x, bias, residual, prob, training) return _bias_dropout_add @torch.jit.script def bias_dropout_add_fused_train(x: torch.Tensor, bias: Optional[torch.Tensor], residual: torch.Tensor, prob: float) -> torch.Tensor: return bias_dropout_add(x, bias, residual, prob, True) @torch.jit.script def bias_dropout_add_fused_inference(x: torch.Tensor, bias: Optional[torch.Tensor], residual: torch.Tensor, prob: float) -> torch.Tensor: return bias_dropout_add(x, bias, residual, prob, False) class ParallelTransformerLayer(MegatronModule): """A single transformer layer. Transformer layer takes input with size [s, b, h] and returns an output of the same size. """ def __init__(self, config, layer_number, layer_type=LayerType.encoder, self_attn_mask_type=AttnMaskType.padding, drop_path_rate=0., num_experts=1): # retriever=None): args = get_args() super(ParallelTransformerLayer, self).__init__() self.layer_number = layer_number self.layer_type = layer_type self.apply_residual_connection_post_layernorm \ = config.apply_residual_connection_post_layernorm self.bf16 = config.bf16 self.fp32_residual_connection = config.fp32_residual_connection # Layernorm on the input data. if args.normalization == 'layernorm': if get_accelerator().device_name() == 'cuda': self.input_layernorm = LayerNorm( config.hidden_size, eps=config.layernorm_epsilon, no_persist_layer_norm=args.no_persist_layer_norm, sequence_parallel=config.sequence_parallel, apply_layernorm_1p=args.apply_layernorm_1p, mem_efficient_ln=args.mem_efficient_ln) else: self.input_layernorm = LayerNorm( config.hidden_size, eps=config.layernorm_epsilon) else: self.input_layernorm = RMSNorm(config.hidden_size, config.layernorm_epsilon) # Self attention. self.self_attention = ParallelAttention( config, layer_number, attention_type=AttnType.self_attn, attn_mask_type=self_attn_mask_type) self.hidden_dropout = config.hidden_dropout self.bias_dropout_fusion = config.bias_dropout_fusion self.drop_path = DropPath(drop_path_rate) if drop_path_rate > 0.0 else None # Layernorm on the attention output if args.normalization == 'layernorm': if get_accelerator().device_name() == 'cuda': self.post_attention_layernorm = LayerNorm( config.hidden_size, eps=config.layernorm_epsilon, no_persist_layer_norm=not config.persist_layer_norm, sequence_parallel=config.sequence_parallel, apply_layernorm_1p=args.apply_layernorm_1p, mem_efficient_ln=args.mem_efficient_ln) else: self.post_attention_layernorm = LayerNorm( config.hidden_size, eps=config.layernorm_epsilon) else: self.post_attention_layernorm = RMSNorm(config.hidden_size, config.layernorm_epsilon) # Cross attention. if self.layer_type in (LayerType.decoder, LayerType.retro_decoder, LayerType.retro_decoder_with_retriever, LayerType.retro_encoder): self.inter_attention = ParallelAttention( config, layer_number, attention_type=AttnType.cross_attn) # Layernorm on the attention output. if args.normalization == 'layernorm': self.post_inter_attention_layernorm = LayerNorm( config.hidden_size, eps=config.layernorm_epsilon, no_persist_layer_norm=not config.persist_layer_norm, sequence_parallel=config.sequence_parallel, apply_layernorm_1p=args.apply_layernorm_1p, mem_efficient_ln=args.mem_efficient_ln) else: self.post_inter_attention_layernorm = RMSNorm(config.hidden_size, config.layernorm_epsilon) # MLP self.num_experts = num_experts if args.num_experts_switch is not None: self.mlp = SwitchMLP(config) # Megatron-LM's MoE else: if self.num_experts <= 1: # dense, not MoE self.mlp = ParallelMLP(config) else: # DeepSpeed's MoE enable_expert_tensor_parallelism = args.enable_expert_tensor_parallelism self.mlp = MoE(args.hidden_size, ParallelMLP(config, moe=True, enable_expert_tensor_parallelism=enable_expert_tensor_parallelism), num_experts=self.num_experts, ep_size=args.moe_expert_parallel_size, k=args.topk, use_residual=(args.mlp_type == 'residual'), capacity_factor=args.moe_train_capacity_factor, eval_capacity_factor=args.moe_eval_capacity_factor, min_capacity=args.moe_min_capacity, drop_tokens=args.moe_token_dropping, use_tutel=args.use_tutel, enable_expert_tensor_parallelism=enable_expert_tensor_parallelism) # Set bias+dropout+add fusion grad_enable execution handler. TORCH_MAJOR = int(torch.__version__.split('.')[0]) TORCH_MINOR = int(torch.__version__.split('.')[1]) use_nvfuser = TORCH_MAJOR > 1 or (TORCH_MAJOR == 1 and TORCH_MINOR >= 10) self.bias_dropout_add_exec_handler = \ nullcontext if use_nvfuser else torch.enable_grad if args.retro_add_retriever: retro_args = get_retro_args() self.retro_num_neighbors = args.retro_num_neighbors self.retro_chunk_length = retro_args.retro_gpt_chunk_length self.retro_retrieved_length = retro_args.retro_gpt_retrieved_length # Retriever (bi-directional transformer with cross attention) if layer_type == LayerType.retro_decoder_with_retriever: self.retriever = ParallelTransformer( init_method, output_layer_init_method, model_type=ModelType.retro_encoder, self_attn_mask_type=AttnMaskType.padding, pre_process=True, post_process=False, ) self._retriever_key = 'retriever' else: self.retriever = None def default_decoder_cross_attention(self, encoder_output, enc_dec_attn_mask, layernorm_input, layernorm_output, bias_dropout_add_func): '''Cross attention for a standard encoder-decoder model.''' # Attention. attention_output, attention_bias = \ self.inter_attention(layernorm_output, enc_dec_attn_mask, encoder_output=encoder_output) # Residual connection. if self.apply_residual_connection_post_layernorm: residual = layernorm_output else: residual = layernorm_input if attention_bias is not None: attention_bias = attention_bias.expand_as(residual) # Bias-dropout-add. with self.bias_dropout_add_exec_handler(): layernorm_input = bias_dropout_add_func( attention_output, attention_bias, residual, self.hidden_dropout) # Layer norm. layernorm_output = self.post_inter_attention_layernorm(layernorm_input) return layernorm_input, layernorm_output def retro_encoder_cross_attention(self, retriever_output, layernorm_input, layernorm_output, bias_dropout_add_func): """Cross attention for Retro encoder. Notation: ns : Sequence length. bs : Batch size. d : Hidden size. l : Number of chunks per sample (i.e., seq_length/chunk_length). k : Number of neighbors. r : Number of retrieved tokens (neighbors + continuation). """ ns, bs, d = layernorm_output.shape # [r, bs * l * k, d] # Divide sequence dimension into chunks. chunked_outputs = layernorm_output.reshape(self.retro_retrieved_length, -1, self.retro_num_neighbors, d) chunked_outputs_before_layer_norm = \ layernorm_input.reshape(self.retro_retrieved_length, -1, self.retro_num_neighbors, d) # [r, bs*l, k, d] # Per-chunk attention. layernorm_inputs = [] layernorm_outputs = [] for k in range(self.retro_num_neighbors): # Attention. chunked_output = chunked_outputs[:,:,k].contiguous() attention_output, attention_bias = \ self.inter_attention( chunked_output, # Q (neighbor embedding) None, encoder_output=retriever_output) # K, V (hidden act) # Residual connection. if self.apply_residual_connection_post_layernorm: residual = chunked_output else: residual = chunked_outputs_before_layer_norm[:,:,k] # Re-enable torch grad to enable fused optimization. with torch.enable_grad(): layernorm_input = bias_dropout_add_func( attention_output, None if attention_bias is None else attention_bias.expand_as(residual), residual, self.hidden_dropout) layernorm_inputs.append(layernorm_input) # Layer norm. layernorm_output = \ self.post_inter_attention_layernorm(layernorm_input) layernorm_outputs.append(layernorm_output) # Concatenate layer norms. # layernorm_input : [r, k * bs * l, d] # layernorm_output : [r, k * bs * l, d] layernorm_input = \ torch.stack(layernorm_inputs, dim=1).reshape(ns, bs, d) layernorm_output = \ torch.stack(layernorm_outputs, dim=1).reshape(ns, bs, d) return layernorm_input, layernorm_output def retro_decoder_cross_attention(self, retriever_input, retriever_output, retriever_attn_mask, layernorm_input, layernorm_output, inference_params, bias_dropout_add_func): """Cross attention for Retro decoder. Notation: ns : Sequence length. bs : Batch size. d : Hidden size. l : Number of chunks per sample (i.e., seq_length/chunk_length). m : Number of tokens per chunk. k : Number of neighbors. r : Number of retrieved tokens (neighbors + continuation). """ ns, bs, d = layernorm_output.shape l = int(np.ceil(ns / self.retro_chunk_length)) # Retrieve neighbors. if self.layer_type == LayerType.retro_decoder_with_retriever: first_ns = ns % self.retro_chunk_length if first_ns > 0: raise Exception("test this case.") first_chunk, rest_chunk = \ layernorm_output[:first_ns], layernorm_output[first_ns:] first_chunk = torch.nn.functional.pad( first_chunk, (0, 0, 0, 0, 0, self.retro_chunk_length - first_ns), 'constant', 0) chunked_output = \ torch.cat((first_chunk, rest_chunk), dim=0) # [l * m, bs, d] else: chunked_output = layernorm_output # [l * m, bs, d] chunked_output = chunked_output \ .reshape(l, self.retro_chunk_length, bs, d) \ .permute(1, 2, 0, 3) \ .reshape(self.retro_chunk_length, bs * l, d) \ .contiguous() # Get Encoder Output retriever_output = self.retriever( hidden_states=retriever_input, attention_mask=retriever_attn_mask, retriever_output=chunked_output, retriever_attn_mask=retriever_attn_mask, inference_params=inference_params) # [r, k * bs * l , d] retriever_output = retriever_output.reshape( self.retro_retrieved_length * self.retro_num_neighbors, bs * l, d) # [r * k, bs * l, d] # Chunks. pad = (ns - 1) % self.retro_chunk_length attending_chunks = layernorm_output[pad:] padded_chunks = torch.nn.functional.pad( attending_chunks, (0, 0, 0, 0, 0, self.retro_chunk_length - 1), 'constant', 0) padded_chunked_output = padded_chunks \ .reshape(l, self.retro_chunk_length, bs, d) \ .permute(1, 2, 0, 3) padded_chunked_output = padded_chunked_output.reshape( self.retro_chunk_length, bs * l, d).contiguous() # Encoder output. attention_output, attention_bias = \ self.inter_attention(padded_chunked_output, None, encoder_output=retriever_output) # Residual connection. if self.apply_residual_connection_post_layernorm: residual = layernorm_output else: residual = layernorm_input # Re-enable torch grad to enable fused optimization. with torch.enable_grad(): layernorm_input = bias_dropout_add_func( attention_output, None if attention_bias is None else attention_bias.expand_as(attention_output), torch.zeros_like(attention_output), self.hidden_dropout) layernorm_input = layernorm_input \ .reshape(self.retro_chunk_length, bs, l, d) \ .permute(2, 0, 1, 3) # [l, m, bs, d] layernorm_input = layernorm_input.reshape(self.retro_chunk_length * l, bs, d) layernorm_input = torch.nn.functional.pad( layernorm_input, (0, 0, 0, 0, pad, 0), 'constant', 0)[:ns] # [ns, b, d] layernorm_input = layernorm_input + residual # Layer norm post the decoder attention layernorm_output = self.post_inter_attention_layernorm(layernorm_input) return retriever_output, layernorm_input, layernorm_output def forward(self, hidden_states, attention_mask=None, encoder_output=None, enc_dec_attn_mask=None, retriever_input=None, retriever_output=None, retriever_attn_mask=None, inference_params=None, rotary_pos_emb=None): # hidden_states: [s, b, h] # Layer norm at the beginning of the transformer layer. layernorm_output = self.input_layernorm(hidden_states) # Self attention. attention_output, attention_bias = \ self.self_attention( layernorm_output, attention_mask, inference_params=inference_params, rotary_pos_emb=rotary_pos_emb) # Residual connection. if self.apply_residual_connection_post_layernorm: residual = layernorm_output else: residual = hidden_states if self.drop_path is None: # jit scripting for a nn.module (with dropout) is not # trigerring the fusion kernel. For now, we use two # different nn.functional routines to account for varying # dropout semantics during training and inference phases. if self.bias_dropout_fusion: if self.training: bias_dropout_add_func = bias_dropout_add_fused_train else: bias_dropout_add_func = bias_dropout_add_fused_inference else: bias_dropout_add_func = get_bias_dropout_add(self.training) if attention_bias is not None: attention_bias = attention_bias.expand_as(residual) with self.bias_dropout_add_exec_handler(): layernorm_input = bias_dropout_add_func( attention_output, attention_bias, residual, self.hidden_dropout) else: out = torch.nn.functional.dropout(attention_output + attention_bias, p=self.hidden_dropout, training=self.training) layernorm_input = residual + self.drop_path(out) # Layer norm post the self attention. layernorm_output = self.post_attention_layernorm(layernorm_input) # Cross attention. if self.layer_type == LayerType.encoder: pass elif self.layer_type == LayerType.decoder: layernorm_input, layernorm_output = \ self.default_decoder_cross_attention( encoder_output, enc_dec_attn_mask, layernorm_input, layernorm_output, bias_dropout_add_func) elif self.layer_type == LayerType.retro_encoder: layernorm_input, layernorm_output = \ self.retro_encoder_cross_attention( retriever_output, layernorm_input, layernorm_output, bias_dropout_add_func) elif self.layer_type in (LayerType.retro_decoder, LayerType.retro_decoder_with_retriever): retriever_output, layernorm_input, layernorm_output = \ self.retro_decoder_cross_attention( retriever_input, retriever_output, retriever_attn_mask, layernorm_input, layernorm_output, inference_params, bias_dropout_add_func) else: raise Exception("Unsupported layer type, '%s'." % self.layer_type.name) # MLP. moe_loss = torch.tensor(0.0, device=layernorm_output.device, dtype=layernorm_output.dtype) mlp_bias = torch.tensor(0.0, device=layernorm_output.device, dtype=layernorm_output.dtype) if self.num_experts == 1: mlp_output, mlp_bias = self.mlp(layernorm_output) else: mlp_output, moe_loss, _ = self.mlp(layernorm_output) # Second residual connection. if self.apply_residual_connection_post_layernorm: residual = layernorm_output else: residual = layernorm_input if self.drop_path is None: if mlp_bias is not None: mlp_bias = mlp_bias.expand_as(residual) with self.bias_dropout_add_exec_handler(): output = bias_dropout_add_func( mlp_output, mlp_bias, residual, self.hidden_dropout) # Jit compiled function creates 'view' tensor. This tensor # potentially gets saved in the MPU checkpoint function context, # which rejects view tensors. While making a viewless tensor here # won't result in memory savings (like the data loader, or # p2p_communication), it serves to document the origin of this # 'view' tensor. output = core.utils.make_viewless_tensor(inp = output, requires_grad = output.requires_grad, keep_graph = True) else: if mlp_bias is not None: mlp_output = mlp_output + mlp_bias out = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training) output = residual + self.drop_path(out) if self.layer_type == LayerType.retro_decoder_with_retriever: return output, retriever_output, moe_loss else: return output, moe_loss class ParallelTransformerLayerPipe(ParallelTransformerLayer): """Extends ParallelTransformerLayer to forward attention_mask through the pipeline. Forward has two usages that affect attention mask communication: 1) forward((input, attn_mask) , **kwargs) -> (output, mask) When the attention mask is provided as the second positional argument, typical pipeline behavior is used and both the output *and* mask are returned in a tuple. This tuple is then forwarded to the next stage in the pipeline. This version is useful if masks are dynamic. 2) forward(input, **kwargs) -> output When the mask is static over all samples, it is advantageous to cache the mask and avoid communicating it. If no mask is provided, the module will query `self._args.attn_mask` for the mask and only return `super().forward(...)` """ def forward(self, inputs, **kwargs): assert torch.is_tensor(inputs) or isinstance(inputs, tuple) if not hasattr(self, '_args'): self._args = get_args() rotary_pos_emb = self._args.rotary_pos_emb if self._args.use_rotary_position_embeddings else None if torch.is_tensor(inputs) or len(inputs) == 1: # No attention mask forwarded, search for args.attn_mask hidden_states, attention_mask = inputs, self._args.attn_mask # HACK: currently MoE model does not support pipeline parallel, so # here we just ignore the moe_loss returned by forward() return super().forward(hidden_states, attention_mask, **kwargs, rotary_pos_emb=rotary_pos_emb)[0] elif len(inputs) == 2: # Attention mask is an activation. hidden_states, attention_mask = inputs[0], inputs[1] # HACK: currently MoE model does not support pipeline parallel, so # here we just ignore the moe_loss returned by forward() return super().forward(*inputs, **kwargs, rotary_pos_emb=rotary_pos_emb)[0], attention_mask else: raise RuntimeError('Received more inputs than understood.') class NoopTransformerLayer(MegatronModule): """A single 'no-op' transformer layer. The sole purpose of this layer is for when a standalone embedding layer is used (i.e., args.standalone_embedding_stage == True). In this case, zero transformer layers are assigned when pipeline rank == 0. Additionally, when virtual pipeline rank >= 1, zero total model parameters are created (virtual rank 0 contains the input embedding). This results in the model's input and output tensors being the same, which causes an error when performing certain memory optimiations on the output tensor (e.g., deallocating it). Thus, this layer disconnects the input from the output via a clone. Since ranks containing a no-op layer are generally under- utilized (both compute and memory), there's no worry of any performance degredation. """ def __init__(self, layer_number): super().__init__() self.layer_number = layer_number def forward(self, hidden_states, attention_mask, encoder_output=None, enc_dec_attn_mask=None, inference_params=None): return hidden_states.clone() def _get_num_layers(args, model_type, is_decoder=False): """Compute the number of transformer layers resident on the current rank.""" is_encoder_and_decoder_model = (model_type == ModelType.encoder_and_decoder) if model_type == ModelType.retro_encoder: num_layers = args.retro_encoder_layers elif parallel_state.get_pipeline_model_parallel_world_size() > 1: if is_encoder_and_decoder_model: assert args.pipeline_model_parallel_split_rank is not None # When a standalone embedding stage is used, a rank is taken from # the encoder's ranks, to be used for the encoder's embedding # layer. This way, the rank referenced by the 'split rank' remains # the same whether or not a standalone embedding stage is used. num_ranks_in_encoder = ( args.pipeline_model_parallel_split_rank - 1 if args.standalone_embedding_stage else args.pipeline_model_parallel_split_rank ) num_ranks_in_decoder = args.transformer_pipeline_model_parallel_size - num_ranks_in_encoder assert args.encoder_num_layers % num_ranks_in_encoder == 0, \ 'encoder_num_layers (%d) must be divisible by number of ranks given to encoder (%d)' % (args.encoder_num_layers, num_ranks_in_encoder) assert args.decoder_num_layers % num_ranks_in_decoder == 0, \ 'decoder_num_layers (%d) must be divisible by number of ranks given to decoder (%d)' % (args.decoder_num_layers, num_ranks_in_decoder) if parallel_state.is_pipeline_stage_before_split(): num_layers = ( 0 if args.standalone_embedding_stage and parallel_state.get_pipeline_model_parallel_rank() == 0 else args.encoder_num_layers // num_ranks_in_encoder ) else: num_layers = args.decoder_num_layers // num_ranks_in_decoder else: assert args.num_layers == args.encoder_num_layers assert args.num_layers % args.transformer_pipeline_model_parallel_size == 0, \ 'num_layers must be divisible by transformer_pipeline_model_parallel_size' # When a standalone embedding stage is used, all transformer layers # are divided among pipeline rank >= 1, while on pipeline rank 0, # ranks either contain the input embedding layer (virtual pp rank 0), # or no layers at all (virtual pp rank >= 1). num_layers = ( 0 if args.standalone_embedding_stage and parallel_state.get_pipeline_model_parallel_rank() == 0 else args.num_layers // args.transformer_pipeline_model_parallel_size ) else: if not is_decoder: num_layers = args.encoder_num_layers else: num_layers = args.decoder_num_layers return num_layers def _get_layer_type(model_type, default_layer_type, retro_layer_numbers, layer_number): args = get_args() if args.retro_add_retriever and layer_number in retro_layer_numbers: if model_type == ModelType.retro_decoder: return LayerType.retro_decoder_with_retriever \ if layer_number == retro_layer_numbers[0] \ else LayerType.retro_decoder elif model_type == ModelType.retro_encoder: return LayerType.retro_encoder else: raise Exception("Unsupported model type, '%s'." % model_type) else: return default_layer_type class ParallelTransformer(MegatronModule): """Transformer class.""" def __init__(self, config, model_type, layer_type=LayerType.encoder, self_attn_mask_type=AttnMaskType.padding, post_layer_norm=True, pre_process=True, post_process=True, drop_path_rate=0.0, num_experts=[1]): super(ParallelTransformer, self).__init__() args = get_args() self.layer_type = layer_type self.model_type = model_type self.bf16 = config.bf16 self.fp32_residual_connection = config.fp32_residual_connection self.post_layer_norm = post_layer_norm self.pre_process = pre_process self.post_process = post_process self.input_tensor = None self.drop_path_rate = drop_path_rate self.transformer_impl = args.transformer_impl self.retro_add_retriever = args.retro_add_retriever self.ds_inference = args.ds_inference # Store activation checkpoiting flag. self.checkpoint_activations = args.checkpoint_activations self.checkpoint_num_layers = args.checkpoint_num_layers self.recompute_granularity = config.recompute_granularity self.recompute_method = config.recompute_method self.recompute_num_layers = config.recompute_num_layers self.distribute_saved_activations = \ config.distribute_saved_activations and not config.sequence_parallel self.sequence_parallel = config.sequence_parallel # Transformer Engine Init. self.transformer_engine_rope_available = False if self.transformer_impl == 'transformer_engine': global transformer_engine import transformer_engine from importlib.metadata import version from pkg_resources import packaging te_version = packaging.version.Version(version("transformer-engine")) if te_version >= packaging.version.Version("0.10.0"): self.transformer_engine_rope_available = True del version, packaging self.use_fp8 = args.fp8_e4m3 or args.fp8_hybrid self.fp8_recipe = None self.fp8_group = None if self.use_fp8: self.fp8_group = parallel_state.get_data_parallel_group() if args.fp8_e4m3: fp8_format = transformer_engine.common.recipe.Format.E4M3 elif args.fp8_hybrid: fp8_format = transformer_engine.common.recipe.Format.HYBRID self.fp8_recipe = transformer_engine.common.recipe.DelayedScaling( margin=args.fp8_margin, interval=args.fp8_interval, fp8_format=fp8_format, amax_history_len=args.fp8_amax_history_len, amax_compute_algo=args.fp8_amax_compute_algo, override_linear_precision=(False, False, not args.fp8_wgrad), ) self.num_microbatches_in_previous_step = -1 self.microbatch_count = 0 self.checkpoint_core_attention = config.recompute_granularity == 'selective' # Number of layers. self.num_layers = _get_num_layers(args, model_type, layer_type==LayerType.decoder) self.drop_path_rates = [ rate.item() for rate in torch.linspace(0, self.drop_path_rate, config.num_layers)] self.retro_layer_numbers = None if model_type == ModelType.retro_decoder: retro_layer_start = 6 if config.num_layers <= 15 else 9 self.retro_layer_numbers = \ np.arange(retro_layer_start, args.num_layers + 1, 3).tolist() if model_type == ModelType.retro_encoder: self.retro_layer_numbers = [1] # Transformer layers. if args.retro_add_retriever: assert self.recompute_granularity != 'full', \ "Full recompute not supported for Retro." assert args.transformer_impl == 'local', \ "Transformer engine does not support Retro layers." def build_layer(layer_number, n_e): if args.transformer_impl == 'local': current_layer_type = _get_layer_type( model_type, layer_type, self.retro_layer_numbers, layer_number) return ParallelTransformerLayer( config, layer_number, layer_type=current_layer_type, self_attn_mask_type=self_attn_mask_type, drop_path_rate=self.drop_path_rates[layer_number - 1], num_experts=n_e) else: assert config.num_attention_heads == config.num_key_value_heads, \ 'Transformer_engine does not support GQA' return transformer_engine.pytorch.TransformerLayer( config.hidden_size, config.ffn_hidden_size, config.num_attention_heads, layernorm_epsilon=config.layernorm_epsilon, hidden_dropout=config.hidden_dropout, attention_dropout=config.attention_dropout, init_method=config.init_method, output_layer_init_method=config.output_layer_init_method, layer_number=layer_number, kv_channels=config.kv_channels, self_attn_mask_type=self_attn_mask_type.name, tp_group=parallel_state.get_tensor_model_parallel_group(), get_rng_state_tracker=tensor_parallel.get_cuda_rng_tracker, fuse_wgrad_accumulation=config.gradient_accumulation_fusion, apply_query_key_layer_scaling=config.apply_query_key_layer_scaling, attention_softmax_in_fp32=config.attention_softmax_in_fp32, seq_length=args.seq_length, micro_batch_size=args.micro_batch_size, sequence_parallel=config.sequence_parallel, params_dtype=config.params_dtype, apply_residual_connection_post_layernorm=config.apply_residual_connection_post_layernorm, output_layernorm=False, layer_type="encoder", drop_path_rate=self.drop_path_rates[layer_number - 1], set_parallel_mode=True, fuse_qkv_params=True) if config.virtual_pipeline_model_parallel_size is not None: assert config.num_layers % config.virtual_pipeline_model_parallel_size == 0, \ 'num_layers_per_stage must be divisible by ' \ 'virtual_pipeline_model_parallel_size' assert args.model_type != ModelType.encoder_and_decoder # Number of layers in each model chunk is the number of layers in the stage, # divided by the number of model chunks in a stage. self.num_layers = self.num_layers // config.virtual_pipeline_model_parallel_size # With 8 layers, 2 stages, and 4 model chunks, we want an assignment of # layers to stages like (each list is a model chunk): # Stage 0: [0] [2] [4] [6] # Stage 1: [1] [3] [5] [7] # With 8 layers, 2 stages, and 2 virtual stages, we want an assignment of # layers to stages like (each list is a model chunk): # Stage 0: [0, 1] [4, 5] # Stage 1: [2, 3] [6, 7] offset = parallel_state.get_virtual_pipeline_model_parallel_rank() * ( config.num_layers // config.virtual_pipeline_model_parallel_size) + \ (parallel_state.get_pipeline_model_parallel_rank() * self.num_layers) else: # Each stage gets a contiguous set of layers. if args.model_type == ModelType.encoder_and_decoder and \ parallel_state.get_pipeline_model_parallel_world_size() > 1: pipeline_rank = parallel_state.get_pipeline_model_parallel_rank() if layer_type == LayerType.encoder: offset = pipeline_rank * self.num_layers else: num_ranks_in_enc = args.pipeline_model_parallel_split_rank offset = (pipeline_rank - num_ranks_in_enc) * self.num_layers else: offset = parallel_state.get_pipeline_model_parallel_rank() * self.num_layers if self.num_layers == 0: # When a standalone embedding stage is used (e.g., # args.standalone_embedding_stage == True), virtual pipeline ranks # on pipeline rank 0 will have zero transformer layers assigned to # them. This results in the model's input and output tensors to be # the same, which will cause failure for certain output tensor # optimizations (e.g., pipeline output deallocation). To remedy # this, we assign a 'no-op' layer on these ranks, which will # disconnect the input tensor from the output tensor. self.num_layers = 1 self.layers = torch.nn.ModuleList([ NoopTransformerLayer(1) ]) else: assert len(num_experts) == 1 or len(num_experts) == args.num_layers // args.expert_interval, \ 'num_experts must be either a single value or a list of the same length as the number of MoE layers' # Create the list of MoE experts if len(num_experts) == 1: num_experts = num_experts * (args.num_layers // args.expert_interval) # Build the layers self.layers = [] for i in range(self.num_layers): layer_num = i + 1 + offset if layer_num % args.expert_interval == 0: n_e = num_experts[(layer_num-1) // args.expert_interval] else: n_e = 1 self.layers.append(build_layer(layer_num, n_e)) self.layers = torch.nn.ModuleList(self.layers) # Update dropout rate for Retro encoder. if model_type == ModelType.retro_encoder: for layer in self.layers: if layer.self_attention.use_flash_attn: layer.self_attention.core_attention_flash.dropout_p = \ torch.nn.Dropout(args.retro_encoder_attention_dropout) else: layer.self_attention.core_attention.attention_dropout.p =\ args.retro_encoder_attention_dropout layer.hidden_dropout = args.retro_encoder_hidden_dropout if self.post_process and self.post_layer_norm: # Final layer norm before output. if args.normalization == 'layernorm': if get_accelerator().device_name() == 'cuda': self.final_layernorm = LayerNorm( config.hidden_size, eps=config.layernorm_epsilon, no_persist_layer_norm=args.no_persist_layer_norm, sequence_parallel=config.sequence_parallel, apply_layernorm_1p=args.apply_layernorm_1p, mem_efficient_ln=args.mem_efficient_ln) else: self.final_layernorm = LayerNorm( config.hidden_size, eps=config.layernorm_epsilon) else: self.final_layernorm = RMSNorm(config.hidden_size, config.layernorm_epsilon) def _get_layer(self, layer_number): return self.layers[layer_number] def _checkpointed_forward(self, hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, rotary_pos_emb, is_first_microbatch): args = get_args() """Forward method with activation checkpointing.""" def custom(start, end): def custom_forward(*args, **kwargs): x_, *args = args moe_losses = [] for index in range(start, end): layer = self._get_layer(index) output = layer(x_, *args, **kwargs) if isinstance(output, tuple): x_, moe_loss = output else: x_ = output moe_loss = torch.tensor(0.0, device=x_.device, dtype=x_.dtype, requires_grad=True) moe_losses.append(moe_loss) return (x_, *moe_losses) return custom_forward if args.deepspeed and args.deepspeed_activation_checkpointing: moe_losses = [] # Make sure memory is freed. tensor_parallel.reset_checkpointed_activations_memory_buffer() l = 0 while l < self.num_layers: hidden_states, *local_moe_losses = tensor_parallel.checkpoint( custom(l, l + self.checkpoint_num_layers), False, hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, None, None, None, None, rotary_pos_emb) moe_losses.extend(local_moe_losses) l += self.checkpoint_num_layers return hidden_states, moe_losses else: moe_losses = [] te_forward_kwargs = {} if self.transformer_impl == 'transformer_engine': te_forward_kwargs['is_first_microbatch'] = is_first_microbatch if self.transformer_engine_rope_available: te_forward_kwargs['rotary_pos_emb'] = rotary_pos_emb if self.recompute_method == 'uniform': # Uniformly divide the total number of Transformer layers and # checkpoint the input activation of each divided chunk. # A method to further reduce memory usage reducing checkpoints. l = 0 while l < self.num_layers: if self.transformer_impl == 'transformer_engine': hidden_states, *local_moe_losses = transformer_engine.pytorch.distributed.checkpoint( custom(l, l + self.recompute_num_layers), self.distribute_saved_activations, tensor_parallel.get_cuda_rng_tracker, mpu.get_tensor_model_parallel_group(), hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, **te_forward_kwargs) else: hidden_states, *local_moe_losses = tensor_parallel.checkpoint( custom(l, l + self.recompute_num_layers), self.distribute_saved_activations, hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, None, None, None, None, rotary_pos_emb) moe_losses.extend(local_moe_losses) l += self.recompute_num_layers elif self.recompute_method == 'block': # Checkpoint the input activation of only a set number of individual # Transformer layers and skip the rest. # A method fully use the device memory removing redundant re-computation. for l in range(self.num_layers): if l < self.recompute_num_layers: if self.transformer_impl == 'transformer_engine': hidden_states, *local_moe_losses = transformer_engine.pytorch.distributed.checkpoint( custom(l, l + 1), self.distribute_saved_activations, tensor_parallel.get_cuda_rng_tracker, mpu.get_tensor_model_parallel_group(), hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, **te_forward_kwargs) else: hidden_states, *local_moe_losses = tensor_parallel.checkpoint( custom(l, l + 1), self.distribute_saved_activations, hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, None, None, None, None, rotary_pos_emb) else: if self.transformer_impl == 'transformer_engine': hidden_states, *local_moe_losses = custom(l, l + 1)( hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, **te_forward_kwargs) else: hidden_states, *local_moe_losses = custom(l, l + 1)( hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, None, None, None, None, rotary_pos_emb) moe_losses.extend(local_moe_losses) else: raise ValueError("Invalid activation recompute method.") return hidden_states, moe_losses def set_input_tensor(self, input_tensor): """Set input tensor to be used instead of forward()'s input. When doing pipeline parallelism the input from the previous stage comes from communication, not from the input, so the model's forward_step_func won't have it. This function is thus used by internal code to bypass the input provided by the forward_step_func""" self.input_tensor = input_tensor def forward(self, hidden_states, attention_mask, encoder_output=None, enc_dec_attn_mask=None, retriever_input=None, retriever_output=None, retriever_attn_mask=None, inference_params=None, rotary_pos_emb=None): # hidden_states: [s, b, h] # Checks. if inference_params: assert self.recompute_granularity is None, \ 'inference does not work with activation checkpointing' # TODO: Below old DeepSpeed code are commented because it's unsure whether # it is still relevant. # # Reza's note: DeepSpeed inference does not support transposes # if not self.ds_inference: # if self.pre_process: # # Data format change to avoid explicit tranposes : [b s h] --> [s b h]. # # If the input flag for fp32 residual connection is set, convert for float. # if self.fp32_residual_connection: # hidden_states = hidden_states.transpose(0, 1).contiguous().float() # # Otherwise, leave it as is. # else: # hidden_states = hidden_states.transpose(0, 1).contiguous() # else: # # See set_input_tensor() # hidden_states = self.input_tensor # if encoder_output is not None: # encoder_output = encoder_output.transpose(0, 1).contiguous() if not self.pre_process: # See set_input_tensor() hidden_states = self.input_tensor # Viewless tensor. # - We only need to create a viewless tensor in the case of micro batch # size (mbs) == 1, since in this case, 'hidden_states.transpose()' # above creates a view tensor, and '.contiguous()' is a pass-through. # For mbs >= 2, '.contiguous()' creates a new tensor, eliminating # the need to make it viewless. # # However, we don't explicitly check mbs == 1 here because # make_viewless_tensor() has negligible overhead when its input # is already viewless. # # - For the 'else' case above, calling make_viewless_tensor() here is # likely redundant, since p2p_communication.py (likely originator) # already creates viewless tensors. That said, make_viewless_tensor() # is called here to be future-proof and corner-case-proof. hidden_states = core.utils.make_viewless_tensor( hidden_states, requires_grad=True, keep_graph=True, ) # RNG context. if self.sequence_parallel: rng_context = tensor_parallel.get_cuda_rng_tracker().fork() else: rng_context = nullcontext() # Forward layers. with rng_context: # The fp8_autocast context manager is a no-op when enabled=True # The if...else serves to short circuit name resolution for fp8_autocast with transformer_engine.pytorch.fp8_autocast( enabled=self.use_fp8, fp8_recipe=self.fp8_recipe, fp8_group=self.fp8_group ) if self.use_fp8 else nullcontext(): # Determine if the current iteration is first microbatch if self.num_microbatches_in_previous_step != get_num_microbatches(): self.microbatch_count = 0 # Reset count on new batch size rampup interval self.num_microbatches_in_previous_step = get_num_microbatches() is_first_microbatch = self.microbatch_count % get_num_microbatches() == 0 # Forward pass. moe_losses = [] if self.checkpoint_activations: hidden_states, moe_losses = self._checkpointed_forward(hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, rotary_pos_emb, is_first_microbatch) elif self.recompute_granularity == 'full': hidden_states, moe_losses = self._checkpointed_forward(hidden_states, attention_mask, encoder_output, enc_dec_attn_mask, rotary_pos_emb, is_first_microbatch) else: forward_kwargs = { 'encoder_output': encoder_output, 'enc_dec_attn_mask': enc_dec_attn_mask, 'inference_params': inference_params, } if self.transformer_impl == 'transformer_engine': forward_kwargs['is_first_microbatch'] = is_first_microbatch forward_kwargs['checkpoint_core_attention'] = self.checkpoint_core_attention if self.transformer_engine_rope_available: forward_kwargs['rotary_pos_emb'] = rotary_pos_emb else: forward_kwargs['rotary_pos_emb'] = rotary_pos_emb forward_kwargs['retriever_input'] = retriever_input forward_kwargs['retriever_output'] = retriever_output forward_kwargs['retriever_attn_mask'] = retriever_attn_mask for index in range(self.num_layers): layer = self._get_layer(index) hidden_states = layer( hidden_states, attention_mask, **forward_kwargs) # First Retro decoder layer returns both hidden_states # and retriever_output. Make retriever_output available # to subsequence Retro layers. if isinstance(hidden_states, tuple): assert (len(hidden_states) == 2 or len(hidden_states) == 3) if len(hidden_states) == 2: if not self.ds_inference: hidden_states, moe_loss = hidden_states moe_losses.append(moe_loss) else: forward_kwargs["retriever_output"] = hidden_states[1] if not self.ds_inference: hidden_states, _, moe_loss = hidden_states moe_losses.append(moe_loss) # Skip counter update for eval and activation checkpointing if torch.is_grad_enabled() and self.training: self.microbatch_count += 1 # Final layer norm. if self.post_process and self.post_layer_norm: # TODO: Below old DeepSpeed code are commented because it's unsure whether # it is still relevant. # if not self.ds_inference: # # Reverting data format change [s b h] --> [b s h]. # hidden_states = hidden_states.transpose(0, 1).contiguous() hidden_states = self.final_layernorm(hidden_states) return (hidden_states, *moe_losses) class LMHeadPipe(MegatronModule): """ Arguments: vocab_size: size of vocabulary. hidden_size: hidden size gather_output: wether output logits being gathered or not. init_method: init method for weight initialization config: """ def __init__(self, hidden_size, vocab_size, config): args = get_args() super(LMHeadPipe, self).__init__() self.lm_head = tensor_parallel.ColumnParallelLinear(input_size=hidden_size, output_size=vocab_size, bias=False, config=config, init_method=config.init_method,) def forward(self, inputs, **kwargs): assert torch.is_tensor(inputs) or isinstance(inputs, tuple) if isinstance(inputs, tuple): hidden_states = inputs[0] else: hidden_states = inputs if not hasattr(self, '_args'): self._args = get_args() if hasattr(self._args, 'attn_mask'): attention_mask = None else: attention_mask = inputs[1] logits, _ = self.lm_head(hidden_states) # If cmd args has attn_mask, we don't forward it as an activation. if hasattr(self._args, 'attn_mask'): return logits else: return logits, attention_mask