add support for Qwen

models/qwen/1_8b_config.json

@@ -0,0 +1,24 @@
+{
+    "emb_size": 2048,
+    "feedforward_size": 5504,
+    "hidden_size": 2048,
+    "hidden_act": "silu",
+    "heads_num": 16,
+    "layers_num": 24,
+    "dropout": 0.0,
+    "data_processor": "lm",
+    "max_seq_length": 8192,
+    "embedding": ["word"],
+    "remove_transformer_bias": true,
+    "remove_attention_bias": false,
+    "remove_embedding_layernorm": true,
+    "rotary_position_embedding": true,
+    "encoder": "transformer",
+    "feed_forward": "gated",
+    "mask": "causal",
+    "layernorm_positioning": "pre",
+    "layernorm": "rms",
+    "target": ["lm"],
+    "use_logn_attn": true,
+    "use_dynamic_ntk": true
+  }

models/qwen/7b_config.json

@@ -0,0 +1,24 @@
+{
+    "emb_size": 4096,
+    "feedforward_size": 11008,
+    "hidden_size": 4096,
+    "hidden_act": "silu",
+    "heads_num": 32,
+    "layers_num": 32,
+    "dropout": 0.0,
+    "data_processor": "lm",
+    "max_seq_length": 8192,
+    "embedding": ["word"],
+    "remove_transformer_bias": true,
+    "remove_attention_bias": false,
+    "remove_embedding_layernorm": true,
+    "rotary_position_embedding": true,
+    "encoder": "transformer",
+    "feed_forward": "gated",
+    "mask": "causal",
+    "layernorm_positioning": "pre",
+    "layernorm": "rms",
+    "target": ["lm"],
+    "use_logn_attn": true,
+    "use_dynamic_ntk": true
+  }

models/qwen_special_tokens_map.json

@@ -0,0 +1,8 @@
+{
+  "pad_token": "<|endoftext|>",
+  "unk_token": "<|UNK|>",
+  "cls_token": "<|im_start|>",
+  "sep_token": "<|im_end|>",
+  "mask_token": "<|MASK|>",
+  "sentinel_token": "<|extra_0|>"
+}

scripts/convert_qwen_from_huggingface_to_tencentpretrain.py

@@ -0,0 +1,54 @@
+import argparse
+import os
+import collections
+from safetensors.torch import load_file
+import torch
+
+parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+parser.add_argument("--input_model_path", type=str, default="models/input_model.bin",
+                        help=".")
+parser.add_argument("--output_model_path", type=str, default="models/output_model.bin",
+                        help=".")
+parser.add_argument("--layers_num", type=int, default=12)
+
+args = parser.parse_args()
+
+input_model = {}
+for file_name in os.listdir(args.input_model_path):
+    if os.path.splitext(file_name)[-1][1:] == "safetensors":
+        dict = load_file(filename=os.path.join(args.input_model_path, file_name))
+        input_model.update(dict)
+
+output_model = collections.OrderedDict()
+emb_size = input_model["transformer.h." + str(0) + ".attn.c_attn.weight"].shape[1]
+
+output_model["embedding.word.embedding.weight"] = input_model["transformer.wte.weight"]
+
+
+for i in range(args.layers_num):
+    for j in range(3):
+        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers." + str(j) + ".weight"] = \
+            input_model["transformer.h." + str(i) + ".attn.c_attn.weight"][j*emb_size:(j+1)*emb_size, :]
+        output_model["encoder.transformer." + str(i) + ".self_attn.linear_layers." + str(j) + ".bias"] = \
+            input_model["transformer.h." + str(i) + ".attn.c_attn.bias"][j*emb_size:(j+1)*emb_size]
+
+    output_model["encoder.transformer." + str(i) + ".self_attn.final_linear.weight"] = \
+        input_model["transformer.h." + str(i) + ".attn.c_proj.weight"]
+
+    output_model["encoder.transformer." + str(i) + ".layer_norm_1.weight"] = \
+        input_model["transformer.h." + str(i) + ".ln_1.weight"]
+
+    output_model["encoder.transformer." + str(i) + ".feed_forward.linear_gate.weight"] = \
+        input_model["transformer.h." + str(i) + ".mlp.w2.weight"]
+    output_model["encoder.transformer." + str(i) + ".feed_forward.linear_1.weight"] = \
+        input_model["transformer.h." + str(i) + ".mlp.w1.weight"]
+    output_model["encoder.transformer." + str(i) + ".feed_forward.linear_2.weight"] = \
+        input_model["transformer.h." + str(i) + ".mlp.c_proj.weight"]
+
+    output_model["encoder.transformer." + str(i) + ".layer_norm_2.weight"] = \
+        input_model["transformer.h." + str(i) + ".ln_2.weight"]
+
+output_model["encoder.layer_norm.weight"] = input_model["transformer.ln_f.weight"]
+output_model["target.lm.output_layer.weight"] = input_model["lm_head.weight"]
+
+torch.save(output_model, args.output_model_path)

scripts/generate_lm.py

@@ -113,6 +113,9 @@ def top_k_top_p_filtering(logits, top_k, top_p):
         if args.tokenizer.sp_model is not None:
             generated_sentence = args.tokenizer.sp_model.decode(tokens)
         else:
-            generated_sentence = "".join(args.tokenizer.convert_ids_to_tokens(tokens))
+            tokens = args.tokenizer.convert_ids_to_tokens(tokens)
+            if hasattr(args.tokenizer, "convert_tokens_to_string"):
+                tokens = args.tokenizer.convert_tokens_to_string(tokens)
+            generated_sentence = "".join(tokens)
 
         f.write(generated_sentence)

tencentpretrain/layers/multi_headed_attn.py

@@ -2,7 +2,7 @@
 import torch
 import torch.nn as nn
 from tencentpretrain import mpu
-from tencentpretrain.utils.rope import apply_rotary_emb
+from tencentpretrain.utils.rope import apply_rotary_emb, apply_rotary_pos_emb
 from tencentpretrain.utils.lora import LoraLinear
 
 
@@ -26,7 +26,7 @@ class MultiHeadedAttention(nn.Module):
     self-attention refers to https://arxiv.org/pdf/1706.03762.pdf
     """
 
-    def __init__(self, hidden_size, heads_num, attention_head_size, local_kv_heads_num, dropout, has_bias=True, with_scale=True,
+    def __init__(self, hidden_size, heads_num, attention_head_size, local_kv_heads_num, dropout, max_seq_length, has_bias=True, has_attention_bias=None, with_scale=True,
                  lora_params=None, layer_number=None):
         super(MultiHeadedAttention, self).__init__()
         self.heads_num = heads_num
@@ -41,6 +41,15 @@ def __init__(self, hidden_size, heads_num, attention_head_size, local_kv_heads_n
         assert heads_num % self.local_kv_heads_num == 0, "heads_num should be divisible by n_local_kv_heads"
         self.repeat_num = self.heads_num // self.local_kv_heads_num
 
+        self.max_seq_length = max_seq_length
+
+        logn_list = [
+            math.log(i, self.max_seq_length) if i > self.max_seq_length else 1
+            for i in range(1, 32768)
+        ]
+        logn_tensor = torch.tensor(logn_list)[None, None, :, None]
+        self.register_buffer("logn_tensor", logn_tensor, persistent=False)
+
         if lora_params is not None:
 
             self.linear_layers = nn.ModuleList(
@@ -53,8 +62,9 @@ def __init__(self, hidden_size, heads_num, attention_head_size, local_kv_heads_n
                              lora_dropout=lora_params['lora_dropout'], bias=has_bias)]
             )
         else:
+            has_attention_bias = has_attention_bias if has_attention_bias is not None else has_bias
             self.linear_layers = nn.ModuleList(
-                [nn.Linear(hidden_size, self.inner_hidden_size, bias=has_bias) if i==0 else nn.Linear(hidden_size, self.kv_embed_dim, bias=has_bias) for i in range(3)]
+                [nn.Linear(hidden_size, self.inner_hidden_size, bias=has_attention_bias) if i==0 else nn.Linear(hidden_size, self.kv_embed_dim, bias=has_attention_bias) for i in range(3)]
             )
         self.dropout = nn.Dropout(dropout)
         self.final_linear = nn.Linear(self.inner_hidden_size, hidden_size, bias=has_bias)
@@ -66,7 +76,7 @@ def __init__(self, hidden_size, heads_num, attention_head_size, local_kv_heads_n
             self.layer_number = None
 
     def forward(self, key, value, query, mask, position_bias=None, has_residual_attention=False, prev_attn=None,
-                freqs_cis=None, alibi=None):
+                freqs_cis=None, alibi=None, use_logn_attn=False, use_dynamic_ntk=False):
         """
         Args:
             key: [batch_size x seq_length x hidden_size]
@@ -103,9 +113,24 @@ def unshape(x):
         key = repeat_kv(key, self.repeat_num).transpose(1, 2)
         value = repeat_kv(value, self.repeat_num).transpose(1, 2)
 
-
         if freqs_cis is not None:
-            query, key = apply_rotary_emb(query.transpose(1,2), key.transpose(1,2), freqs_cis=freqs_cis)
+            if use_dynamic_ntk:
+                rotary_pos_emb = freqs_cis
+                rotary_pos_emb = [i[:, -seq_length:, :, :] for i in rotary_pos_emb]
+                rotary_pos_emb = (rotary_pos_emb,) * 2
+                q_pos_emb, k_pos_emb = rotary_pos_emb
+                # Slice the pos emb for current inference
+                query = apply_rotary_pos_emb(query.transpose(1,2), q_pos_emb)
+                key = apply_rotary_pos_emb(key.transpose(1,2), k_pos_emb)
+            else:
+                query, key = apply_rotary_emb(query.transpose(1,2), key.transpose(1,2), freqs_cis=freqs_cis)
+
+        key_size = key.size(2)
+        if key_size > self.max_seq_length and use_logn_attn and not self.training:
+            seq_start = key_size - query.size(2)
+            seq_end = key_size
+            logn_tensor = self.logn_tensor[:, :, seq_start:seq_end, :].type_as(query)
+            query = query * logn_tensor.expand_as(query)
 
         scores = torch.matmul(query, key.transpose(-2, -1))
 

tencentpretrain/layers/transformer.py

@@ -1,5 +1,6 @@
 import torch
 import torch.nn as nn
+from tencentpretrain.utils.rope import get_ntk_alpha, update_freqs_cis
 from tencentpretrain.layers.multi_headed_attn import MultiHeadedAttention, ParallelMultiHeadedAttention
 from tencentpretrain.layers import *
 
@@ -16,6 +17,13 @@ def __init__(self, args, layer_number=None):
         self.relative_position_embedding = args.relative_position_embedding
         self.rotary_position_embedding = args.rotary_position_embedding
         self.has_residual_attention = args.has_residual_attention
+        self.use_logn_attn = args.use_logn_attn
+        self.max_seq_length = args.max_seq_length
+        self.use_dynamic_ntk = args.use_dynamic_ntk
+        self.hidden_size = args.hidden_size
+        self.heads_num = args.heads_num
+        self.seq_len_cached = 0
+        self.ntk_alpha_cached = 1.0
         if self.relative_position_embedding:
             self.relative_pos_emb = args.relative_pos_emb
         if self.rotary_position_embedding:
@@ -32,6 +40,7 @@ def __init__(self, args, layer_number=None):
             local_kv_heads_num = args.heads_num
 
         has_bias = bool(1 - args.remove_transformer_bias)
+        has_attention_bias = bool(1 - args.remove_attention_bias) if hasattr(args, "remove_attention_bias") else None
         with_scale = bool(1 - args.remove_attention_scale)
 
         # Multi-headed self-attention.
@@ -40,8 +49,8 @@ def __init__(self, args, layer_number=None):
             lora_params = args.lora_params
 
         self.self_attn = MultiHeadedAttention(
-            args.hidden_size, args.heads_num, attention_head_size, local_kv_heads_num, args.dropout, has_bias=has_bias,
-            with_scale=with_scale, lora_params=lora_params, layer_number=layer_number
+            args.hidden_size, args.heads_num, attention_head_size, local_kv_heads_num, args.dropout, self.max_seq_length, has_bias=has_bias, has_attention_bias = has_attention_bias,
+            with_scale = with_scale, lora_params=lora_params, layer_number=layer_number
         )
         self.dropout_1 = nn.Dropout(args.dropout)
 
@@ -77,22 +86,32 @@ def forward(self, *inputs):
         else:
             position_bias = None
 
-        if self.rotary_position_embedding:
+        if self.use_dynamic_ntk:
+            ntk_alpha = get_ntk_alpha(seq_length, self.max_seq_length) if seq_length > self.max_seq_length else 1.0
+            if seq_length > self.seq_len_cached or ntk_alpha != self.ntk_alpha_cached:
+                self.freqs_cis = update_freqs_cis(self.hidden_size // self.heads_num, seq_length * 2, ntk_alpha=ntk_alpha)
+                self.seq_len_cached = seq_length * 2
+                self.ntk_alpha_cached = ntk_alpha
+
+        if self.rotary_position_embedding and not self.use_dynamic_ntk:
             freqs_cis = self.freqs_cis[:seq_length].to(hidden.device)
+        elif self.use_dynamic_ntk:
+            cos, sin = self.freqs_cis
+            freqs_cis = [cos[:, :seq_length], sin[:, :seq_length]]
         else:
             freqs_cis = None
 
         if self.layernorm_positioning == "post":
             inter, prev_attn_out = self.self_attn(hidden, hidden, hidden, mask, position_bias, self.has_residual_attention,
-                                                  prev_attn, freqs_cis)
+                                                  prev_attn, freqs_cis, use_logn_attn=self.use_logn_attn, use_dynamic_ntk=self.use_dynamic_ntk)
             inter = self.dropout_1(inter)
             inter = self.layer_norm_1(inter + hidden)
             output = self.dropout_2(self.feed_forward(inter))
             output = self.layer_norm_2(output + inter)
         else:
             inter = self.layer_norm_1(hidden)
             inter, prev_attn_out = self.self_attn(inter, inter, inter, mask, position_bias, self.has_residual_attention,
-                                                  prev_attn, freqs_cis)
+                                                  prev_attn, freqs_cis, use_logn_attn=self.use_logn_attn, use_dynamic_ntk=self.use_dynamic_ntk)
             inter = self.dropout_1(inter)
             hidden = hidden + inter
             output = self.layer_norm_2(hidden)
@@ -281,14 +300,14 @@ def __init__(self, args):
             lora_params = args.lora_params
 
         self.self_attn = MultiHeadedAttention(
-            args.hidden_size, args.heads_num, attention_head_size, local_kv_heads_num, args.dropout, has_bias=has_bias,
+            args.hidden_size, args.heads_num, attention_head_size, local_kv_heads_num, args.dropout, args.max_seq_length, has_bias=has_bias,
             with_scale=with_scale, lora_params=lora_params
         )
         self.dropout_1 = nn.Dropout(args.dropout)
 
         # Multi-headed context-attention.
         self.context_attn = MultiHeadedAttention(
-            args.hidden_size, args.heads_num, attention_head_size, local_kv_heads_num, args.dropout, has_bias=has_bias,
+            args.hidden_size, args.heads_num, attention_head_size, local_kv_heads_num, args.dropout, args.max_seq_length, has_bias=has_bias,
             with_scale=with_scale, lora_params=lora_params
         )
         self.dropout_2 = nn.Dropout(args.dropout)

tencentpretrain/opts.py

@@ -176,7 +176,7 @@ def infer_opts(parser):
 
 def tokenizer_opts(parser):
     parser.add_argument("--tokenizer", choices=["bert", "bpe", "char", "space", "xlmroberta", "image", "text_image",
-                                                "virtual", "hfpretrained"], default="bert",
+                                                "virtual", "hfpretrained", "qwen"], default="bert",
                         help="Specify the tokenizer." 
                              "Original Google BERT uses bert tokenizer."
                              "Char tokenizer segments sentences into characters."

tencentpretrain/utils/__init__.py

@@ -7,7 +7,8 @@
 
 str2tokenizer = {"char": CharTokenizer, "space": SpaceTokenizer, "bert": BertTokenizer,
                  "bpe": BPETokenizer, "xlmroberta": XLMRobertaTokenizer, "image": ImageTokenizer,
-                 "text_image": TextImageTokenizer, "virtual": VirtualTokenizer, "hfpretrained": HFPreTrainedTokenizer}
+                 "text_image": TextImageTokenizer, "virtual": VirtualTokenizer, "hfpretrained": HFPreTrainedTokenizer,
+                 "qwen": QwenTokenizer}
 str2dataset = {"bert": BertDataset, "lm": LmDataset, "mlm": MlmDataset,
                "bilm": BilmDataset, "albert": AlbertDataset, "mt": MtDataset,
                "t5": T5Dataset, "gsg": GsgDataset, "bart": BartDataset,

tencentpretrain/utils/rope.py

@@ -1,4 +1,5 @@
 import torch
+import math
 from typing import Tuple
 
 def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
@@ -8,6 +9,24 @@ def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
     freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
     return freqs_cis
 
+def get_ntk_alpha(true_seq_len, max_seq_length):
+    context_value = math.log(true_seq_len / max_seq_length, 2) + 1
+    ntk_alpha = 2 ** math.ceil(context_value) - 1
+    ntk_alpha = max(ntk_alpha, 1)
+    return ntk_alpha
+
+def update_freqs_cis(dim: int, end: int, theta: float = 10000.0, ntk_alpha: float = 1.0):
+    theta = theta * ntk_alpha ** (dim / (dim - 2))
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2).float() / dim))
+    seq = torch.arange(end, device=freqs.device)
+    freqs = torch.outer(seq.type_as(freqs), freqs)
+    emb = torch.cat((freqs, freqs), dim=-1)
+
+    from einops import rearrange
+    emb = rearrange(emb, "n d -> 1 n 1 d")
+
+    cos, sin = emb.cos(), emb.sin()
+    return [cos[:, :end], sin[:, :end]]
 
 def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
     ndim = x.ndim
@@ -28,3 +47,20 @@ def apply_rotary_emb(
     xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
     xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
     return xq_out.type_as(xq).transpose(1,2), xk_out.type_as(xk).transpose(1,2)
+
+def _rotate_half(x: torch.Tensor):
+    from einops import rearrange
+
+    x = rearrange(x, "... (j d) -> ... j d", j=2)
+    x1, x2 = x.unbind(dim=-2)
+    return torch.cat((-x2, x1), dim=-1)
+
+def apply_rotary_pos_emb(t: torch.Tensor, freqs: list[torch.Tensor, torch.Tensor]):
+    rot_dim = freqs[0].shape[-1]
+    cos, sin = freqs
+    t_float = t.float()
+    t_rot, t_pass = t_float[..., :rot_dim], t_float[..., rot_dim:]
+    cos = cos.to(t_rot.device)
+    sin = sin.to(t_rot.device)
+    t_rot = (t_rot * cos) + (_rotate_half(t_rot) * sin)
+    return torch.cat((t_rot, t_pass), dim=-1).type_as(t).transpose(1, 2)