import time from typing import Dict, Optional import torch from ...quantization import QuantAlgo from ..convert_utils import (get_weight, get_weight_and_bias, iterate_shard_files, load_state_dict, retrieved_layer_index_from_name) from .config import FalconConfig def split(weight: torch.Tensor, tp_size: int, rank: int = 0, dim: int = 0) -> torch.Tensor: if tp_size == 1: return weight elif weight.ndim == 1: return torch.chunk(weight, tp_size)[rank].clone() else: return torch.chunk(weight, tp_size, dim=dim)[rank].clone() def reorder_qkv_weight_or_bias(weight: torch.Tensor, head_dim: int, num_heads: int, num_kv_heads: Optional[int] = None, tp_size: int = 1, is_bias: bool = False) -> torch.Tensor: """ Reorder the qkv weight for TRT-LLM use. The shape of the fused QKV weights in HF is different from the shape that TRT-LLM requires. In particular, the weight of HF consists of interleaved q, k, v head weights, while that of TRT-LLM is contiguous. HF : [q1, k1, v1, ..., qh, kh, vh] TRT-LLM: [q1, ..., qh, k1, ..., kh, v1, vh] where qi, vi, ki are weight vectors corresponding to attention head i. It's similar to multi/grouped query attention cases. We reorder and split the weight of an attention layer to fit into TRT-LLM. The reordered weight and bias will be weight: (T, Qh * D + 2 * KVh * D, H) bias : (T, Qh * D + 2 * KVh * D) where T=tp_size, Qh=local_num_q_heads, KVh=local_num_kv_heads, D=head_dim, H=hidden_dim. In the multi/grouped query attention, the number of K/V attention heads are less than that of Q attention, so that K/V attention heads may be shared across different ranks if necessary. For tensor parallelism, we use the first dimension to select the corresponding weights. """ # Query types and expected kv heads. # - Conventional MHA: num_heads = num_kv_heads # - Multi-Query Attention: num_kv_heads = 1 # - Grouped-Query Attention: num_heads % num_kv_heads = 0 num_kv_heads = num_kv_heads if num_kv_heads is not None else num_heads assert num_heads % num_kv_heads == 0, \ f'num_heads({num_heads}) must be divisible by ' \ f'num_kv_heads({num_kv_heads})).' # The number of attention heads per group: N q head + 1 k head + 1 v head. num_group_heads = num_heads // num_kv_heads + 2 assert weight.shape[0] == num_kv_heads * num_group_heads * head_dim, \ f'{weight.shape[0]} != {num_kv_heads} * {num_group_heads} * {head_dim}' qkv_in = num_heads * head_dim if not is_bias else 1 # Split Q/K/V weights weight = weight.reshape(num_kv_heads, num_heads // num_kv_heads + 2, head_dim, qkv_in) q_w = weight[:, :-2, ...] # (nKV, num_heads // nKV, head_dim, qkv_in) k_w = weight[:, -2:-1, ...] # (nKV, 1, head_dim, qkv_in) v_w = weight[:, -1:, ...] # (nKV, 1, head_dim, qkv_in) if num_kv_heads < num_heads and num_kv_heads < tp_size: # Duplicate K/V heads to make sure that each rank has at least one # K/V heads. For instance, num_heads=8, num_kv_heads=2, tp_size=4, # we will make the qkv weight as below. # Orig: [q0 q1 q2 q3 k0 v0 q4 q5 q6 q7 k1 v0 v1] # >>>> [[q0 q1 k0 v0], [q2 q3 k0 v0], [q4 q5 k1 v1], [q6 q7 k1 v1]] assert tp_size % num_kv_heads == 0 num_dups = tp_size // num_kv_heads # k_w and v_w have the same shape. new_shape = (num_kv_heads, num_dups) + k_w.shape[2:] k_w = torch.broadcast_to(k_w, size=new_shape) v_w = torch.broadcast_to(v_w, size=new_shape) # Update the number of kv heads. num_kv_heads = tp_size reordered = torch.concat( [ q_w.reshape(tp_size, num_heads // tp_size, head_dim, qkv_in), k_w.reshape(tp_size, num_kv_heads // tp_size, head_dim, qkv_in), v_w.reshape(tp_size, num_kv_heads // tp_size, head_dim, qkv_in), ], dim=1, ) qkv_out = (num_heads + 2 * num_kv_heads) // tp_size * head_dim return reordered.reshape((tp_size, qkv_out, -1)) def split_qkv_weight(weight: torch.Tensor, hidden_size: int, num_heads: int, tp_size: int, rank: int, is_bias: bool, num_kv_heads: Optional[int] = None) -> torch.Tensor: """ Splits the QKV matrix according to tensor parallelism """ head_dim = hidden_size // num_heads weight = reorder_qkv_weight_or_bias(weight, head_dim=head_dim, num_heads=num_heads, num_kv_heads=num_kv_heads, tp_size=tp_size, is_bias=is_bias) # Copy a sliced tensor to prevent memory leak. A sliced tensor shares the # memory buffer of the original tensor. So, returning without copying makes # the buffer of a loaded "qkv" be referenced, resulting GC can't release # those weights until the whole process ends. if not is_bias: return weight[rank, ...].clone() else: return weight[rank, ...].ravel().clone() def split_matrix(weight: torch.Tensor, tp_size: int, rank: int, dim: int) -> torch.Tensor: return split(weight, tp_size, rank, dim=dim) def get_tllm_linear_weight( weight: torch.Tensor, prefix: str, bias: Optional[torch.Tensor] = None, use_weight_only: bool = False, plugin_weight_only_quant_type: torch.dtype = torch.int8 ) -> Dict[str, torch.Tensor]: results = {} if use_weight_only: v = weight.t().contiguous() processed_torch_weights, torch_weight_scales = \ torch.ops.trtllm.symmetric_quantize_last_axis_of_batched_matrix( v, plugin_weight_only_quant_type) results[f'{prefix}.weight'] = processed_torch_weights results[f'{prefix}.per_channel_scale'] = torch_weight_scales else: results[f'{prefix}.weight'] = weight if bias is not None: results[f'{prefix}.bias'] = bias return results def get_tllm_param( param: torch.Tensor, name: str, use_weight_only: bool = False, plugin_weight_only_quant_type: torch.dtype = torch.int8 ) -> Dict[str, torch.Tensor]: results = {} if name.endswith('.weight') and use_weight_only: v = param.t().contiguous() processed_torch_weights, torch_weight_scales = \ torch.ops.trtllm.symmetric_quantize_last_axis_of_batched_matrix( v, plugin_weight_only_quant_type) results[name] = processed_torch_weights results[name.replace('weight', 'per_channel_scale')] = torch_weight_scales else: results[name] = param return results def load_weights_from_hf_model(model, config: FalconConfig): weights = {} tik = time.time() model_params = dict(model.named_parameters()) dtype = getattr(torch, config.dtype) mapping = config.mapping num_attention_heads = config.num_attention_heads hidden_size = config.hidden_size vocab_size = config.vocab_size num_kv_heads = getattr(config, 'num_key_value_heads', num_attention_heads) num_hidden_layers = config.num_hidden_layers parallel_attention = config.parallel_attention new_decoder_architecture = config.new_decoder_architecture use_parallel_embedding = config.use_parallel_embedding sharding_dim = config.embedding_sharding_dim quant_algo = config.quantization.quant_algo use_weight_only = quant_algo in [QuantAlgo.W8A16, QuantAlgo.W4A16] if quant_algo == QuantAlgo.W8A16: plugin_weight_only_quant_type = torch.int8 elif quant_algo == QuantAlgo.W4A16: plugin_weight_only_quant_type = torch.quint4x2 else: plugin_weight_only_quant_type = None layers_range = mapping.pp_layers(num_hidden_layers) for l in layers_range: prefix = f'transformer.h.{l}' tllm_prex = f'transformer.layers.{l - layers_range[0]}' qkv_weight, qkv_bias = get_weight_and_bias( model_params, f'{prefix}.self_attention.query_key_value', dtype) qkv_w = split_qkv_weight(qkv_weight, hidden_size, num_attention_heads, mapping.tp_size, mapping.tp_rank, is_bias=False, num_kv_heads=num_kv_heads) if qkv_bias is None: qkv_b = None else: qkv_b = split_qkv_weight(qkv_bias, hidden_size, num_attention_heads, mapping.tp_size, mapping.tp_rank, is_bias=True, num_kv_heads=num_kv_heads) weights.update( get_tllm_linear_weight(qkv_w, f'{tllm_prex}.attention.qkv', qkv_b, use_weight_only, plugin_weight_only_quant_type)) attn_dense_weight, attn_dense_bias = get_weight_and_bias( model_params, f'{prefix}.self_attention.dense', dtype) attn_dense_w = split_matrix(attn_dense_weight, mapping.tp_size, mapping.tp_rank, dim=1) weights.update( get_tllm_linear_weight(attn_dense_w, f'{tllm_prex}.attention.dense', attn_dense_bias, use_weight_only, plugin_weight_only_quant_type)) mlp_fc_weight, mlp_fc_bias = get_weight_and_bias( model_params, f'{prefix}.mlp.dense_h_to_4h', dtype) mlp_fc_w = split_matrix(mlp_fc_weight, mapping.tp_size, mapping.tp_rank, dim=0) if mlp_fc_bias is None: mlp_fc_b = None else: mlp_fc_b = split_matrix(mlp_fc_bias, mapping.tp_size, mapping.tp_rank, dim=0) weights.update( get_tllm_linear_weight(mlp_fc_w, f'{tllm_prex}.mlp.fc', mlp_fc_b, use_weight_only, plugin_weight_only_quant_type)) mlp_proj_weight, mlp_proj_bias = get_weight_and_bias( model_params, f'{prefix}.mlp.dense_4h_to_h', dtype) mlp_proj_w = split_matrix(mlp_proj_weight, mapping.tp_size, mapping.tp_rank, dim=1) weights.update( get_tllm_linear_weight(mlp_proj_w, f'{tllm_prex}.mlp.proj', mlp_proj_bias, use_weight_only, plugin_weight_only_quant_type)) if new_decoder_architecture: input_ln_weight, input_ln_bias = get_weight_and_bias( model_params, f'{prefix}.ln_attn', dtype) if input_ln_weight is None: input_ln_weight, input_ln_bias = get_weight_and_bias( model_params, f'{prefix}.input_layernorm', dtype) weights[f'{tllm_prex}.input_layernorm.weight'] = input_ln_weight if input_ln_bias is not None: weights[f'{tllm_prex}.input_layernorm.bias'] = input_ln_bias mlp_ln_weight, mlp_ln_bias = get_weight_and_bias( model_params, f'{prefix}.ln_mlp', dtype) if mlp_ln_weight is not None: weights[f'{tllm_prex}.mlp_layernorm.weight'] = mlp_ln_weight if mlp_ln_bias is not None: weights[f'{tllm_prex}.mlp_layernorm.bias'] = mlp_ln_bias else: input_ln_weight, input_ln_bias = get_weight_and_bias( model_params, f'{prefix}.input_layernorm', dtype) weights[f'{tllm_prex}.input_layernorm.weight'] = input_ln_weight if input_ln_bias is not None: weights[f'{tllm_prex}.input_layernorm.bias'] = input_ln_bias if not parallel_attention: post_ln_weight, post_ln_bias = get_weight_and_bias( model_params, f'{prefix}.post_attention_layernorm', dtype) if post_ln_weight is not None: weights[ f'{tllm_prex}.post_layernorm.weight'] = post_ln_weight if post_ln_bias is not None: weights[f'{tllm_prex}.post_layernorm.bias'] = post_ln_bias embed_w = get_weight(model_params, 'transformer.word_embeddings', dtype) if mapping.is_first_pp_rank(): if not use_parallel_embedding: weights['transformer.vocab_embedding.weight'] = embed_w else: if sharding_dim == 0: assert vocab_size % mapping.tp_size == 0 else: assert hidden_size % mapping.tp_size == 0 weights['transformer.vocab_embedding.weight'] = split_matrix( embed_w, mapping.tp_size, mapping.tp_rank, sharding_dim) if mapping.is_last_pp_rank(): lm_head = get_weight(model_params, 'lm_head', dtype) if lm_head is None: # No lm_head in the checkpoint, cloning word_embedding. lm_head = embed_w.clone() weights['lm_head.weight'] = split_matrix(lm_head, mapping.tp_size, mapping.tp_rank, dim=0) ln_f_w, ln_f_b = get_weight_and_bias(model_params, 'transformer.ln_f', dtype) weights['transformer.ln_f.weight'] = ln_f_w if ln_f_b is not None: weights['transformer.ln_f.bias'] = ln_f_b tok = time.time() t = time.strftime('%H:%M:%S', time.gmtime(tok - tik)) print(f'Weights loaded. Total time: {t}') return weights def load_weights_from_hf_by_shard(model_dir: str, config: FalconConfig): weights = {} tik = time.time() dtype = getattr(torch, config.dtype) mapping = config.mapping num_attention_heads = config.num_attention_heads hidden_size = config.hidden_size vocab_size = config.vocab_size num_kv_heads = getattr(config, 'num_key_value_heads', num_attention_heads) num_hidden_layers = config.num_hidden_layers use_weight_only = config.quantization.quant_algo in [ QuantAlgo.W8A16, QuantAlgo.W4A16 ] if config.quantization.quant_algo == QuantAlgo.W8A16: plugin_weight_only_quant_type = torch.int8 elif config.quantization.quant_algo == QuantAlgo.W4A16: plugin_weight_only_quant_type = torch.quint4x2 else: plugin_weight_only_quant_type = None layers_range = mapping.pp_layers(num_hidden_layers) for model_file in iterate_shard_files(model_dir, mapping.tp_rank): state_dict = load_state_dict(model_file, dtype) for name, param in state_dict.items(): l = retrieved_layer_index_from_name(name) if l is not None: if l not in layers_range: continue prefix = f'transformer.layers.{l-layers_range[0]}' if 'self_attention.query_key_value' in name: if name.endswith('weight'): qkv_w = split_qkv_weight(param, hidden_size, num_attention_heads, mapping.tp_size, mapping.tp_rank, is_bias=False, num_kv_heads=num_kv_heads) weights.update( get_tllm_param(qkv_w, f'{prefix}.attention.qkv.weight', use_weight_only, plugin_weight_only_quant_type)) else: qkv_b = split_qkv_weight(param, hidden_size, num_attention_heads, mapping.tp_size, mapping.tp_rank, is_bias=True, num_kv_heads=num_kv_heads) weights.update( get_tllm_param(qkv_b, f'{prefix}.attention.qkv.bias', use_weight_only, plugin_weight_only_quant_type)) elif 'self_attention.dense' in name: if name.endswith('weight'): attn_dense_w = split_matrix(param, mapping.tp_size, mapping.tp_rank, dim=1) weights.update( get_tllm_param(attn_dense_w, f'{prefix}.attention.dense.weight', use_weight_only, plugin_weight_only_quant_type)) else: weights.update( get_tllm_param(param, f'{prefix}.attention.dense.bias', use_weight_only, plugin_weight_only_quant_type)) elif 'mlp.dense_h_to_4h' in name: if name.endswith('weight'): mlp_fc_w = split_matrix(param, mapping.tp_size, mapping.tp_rank, dim=0) weights.update( get_tllm_param(mlp_fc_w, f'{prefix}.mlp.fc.weight', use_weight_only, plugin_weight_only_quant_type)) else: mlp_fc_b = split_matrix(param, mapping.tp_size, mapping.tp_rank, dim=0) weights.update( get_tllm_param(mlp_fc_b, f'{prefix}.mlp.fc.bias', use_weight_only, plugin_weight_only_quant_type)) elif 'mlp.dense_4h_to_h' in name: if name.endswith('weight'): mlp_proj_w = split_matrix(param, mapping.tp_size, mapping.tp_rank, dim=1) weights.update( get_tllm_param(mlp_proj_w, f'{prefix}.mlp.proj.weight', use_weight_only, plugin_weight_only_quant_type)) else: weights.update( get_tllm_param(param, f'{prefix}.mlp.proj.bias', use_weight_only, plugin_weight_only_quant_type)) elif 'ln_attn' in name or 'input_layernorm' in name: if name.endswith('weight'): weights[f'{prefix}.input_layernorm.weight'] = param else: weights[f'{prefix}.input_layernorm.bias'] = param elif 'ln_mlp' in name: if name.endswith('weight'): weights[f'{prefix}.mlp_layernorm.weight'] = param else: weights[f'{prefix}.mlp_layernorm.bias'] = param elif 'post_attention_layernorm' in name: if name.endswith('weight'): weights[f'{prefix}.post_layernorm.weight'] = param else: weights[f'{prefix}.post_layernorm.bias'] = param elif 'word_embeddings' in name: if mapping.is_first_pp_rank(): if not config.use_parallel_embedding: weights['transformer.vocab_embedding.weight'] = param else: if config.embedding_sharding_dim == 0: assert vocab_size % mapping.tp_size == 0 else: assert hidden_size % mapping.tp_size == 0 weights[ 'transformer.vocab_embedding.weight'] = split_matrix( param, mapping.tp_size, mapping.tp_rank, config.embedding_sharding_dim) if mapping.is_last_pp_rank(): weights['lm_head.weight'] = split_matrix(param, mapping.tp_size, mapping.tp_rank, dim=0) elif 'ln_f' in name: if mapping.is_last_pp_rank(): if name.endswith('weight'): weights['transformer.ln_f.weight'] = param else: weights['transformer.ln_f.bias'] = param del state_dict tok = time.time() t = time.strftime('%H:%M:%S', time.gmtime(tok - tik)) print(f'Weights loaded. Total time: {t}') return weights