Update implementation

config.json

@@ -35,7 +35,7 @@
   "seq_length": 8192,
   "use_cache": true,
   "torch_dtype": "bfloat16",
-  "transformers_version": "4.27.1",
+  "transformers_version": "4.30.2",
   "tie_word_embeddings": false,
   "eos_token_id": 2
 }

configuration_chatglm.py

@@ -20,18 +20,19 @@ class ChatGLMConfig(PretrainedConfig):
         post_layer_norm=True,
         add_bias_linear=False,
         add_qkv_bias=False,
-        interleaved_qkv=False,
         bias_dropout_fusion=True,
-        rotary_percent=1.0,
         multi_query_attention=False,
         multi_query_group_num=1,
         apply_query_key_layer_scaling=True,
         attention_softmax_in_fp32=True,
         fp32_residual_connection=False,
         quantization_bit=0,
+        pre_seq_len=None,
+        prefix_projection=False,
         **kwargs
     ):
         self.num_layers = num_layers
+        self.vocab_size = padded_vocab_size
         self.padded_vocab_size = padded_vocab_size
         self.hidden_size = hidden_size
         self.ffn_hidden_size = ffn_hidden_size
@@ -46,13 +47,13 @@ class ChatGLMConfig(PretrainedConfig):
         self.post_layer_norm = post_layer_norm
         self.add_bias_linear = add_bias_linear
         self.add_qkv_bias = add_qkv_bias
-        self.interleaved_qkv = interleaved_qkv
         self.bias_dropout_fusion = bias_dropout_fusion
-        self.rotary_percent = rotary_percent
         self.multi_query_attention = multi_query_attention
         self.multi_query_group_num = multi_query_group_num
         self.apply_query_key_layer_scaling = apply_query_key_layer_scaling
         self.attention_softmax_in_fp32 = attention_softmax_in_fp32
         self.fp32_residual_connection = fp32_residual_connection
         self.quantization_bit = quantization_bit
-        super().__init__(**kwargs)
+        self.pre_seq_len = pre_seq_len
+        self.prefix_projection = prefix_projection
+        super().__init__(**kwargs)

generation_config.json

@@ -0,0 +1,5 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 2,
+  "transformers_version": "4.30.2"
+}

modeling_chatglm.py

@@ -35,12 +35,12 @@ if sys.platform != 'darwin':
 
 logger = logging.get_logger(__name__)
 
-_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM-6B"
+_CHECKPOINT_FOR_DOC = "THUDM/ChatGLM2-6B"
 _CONFIG_FOR_DOC = "ChatGLM6BConfig"
 
 CHATGLM_6B_PRETRAINED_MODEL_ARCHIVE_LIST = [
-    "THUDM/chatglm-6b",
-    # See all ChatGLM-6B models at https://huggingface.co/models?filter=chatglm
+    "THUDM/chatglm2-6b",
+    # See all ChatGLM models at https://huggingface.co/models?filter=chatglm
 ]
 
 
@@ -56,6 +56,38 @@ class InvalidScoreLogitsProcessor(LogitsProcessor):
         return scores
 
 
+class PrefixEncoder(torch.nn.Module):
+    """
+    The torch.nn model to encode the prefix
+    Input shape: (batch-size, prefix-length)
+    Output shape: (batch-size, prefix-length, 2*layers*hidden)
+    """
+
+    def __init__(self, config: ChatGLMConfig):
+        super().__init__()
+        self.prefix_projection = config.prefix_projection
+        if self.prefix_projection:
+            # Use a two-layer MLP to encode the prefix
+            kv_size = config.num_layers * config.kv_channels * config.multi_query_group_num * 2
+            self.embedding = torch.nn.Embedding(config.pre_seq_len, kv_size)
+            self.trans = torch.nn.Sequential(
+                torch.nn.Linear(kv_size, config.hidden_size),
+                torch.nn.Tanh(),
+                torch.nn.Linear(config.hidden_size, kv_size)
+            )
+        else:
+            self.embedding = torch.nn.Embedding(config.pre_seq_len,
+                                                config.num_layers * config.kv_channels * config.multi_query_group_num * 2)
+
+    def forward(self, prefix: torch.Tensor):
+        if self.prefix_projection:
+            prefix_tokens = self.embedding(prefix)
+            past_key_values = self.trans(prefix_tokens)
+        else:
+            past_key_values = self.embedding(prefix)
+        return past_key_values
+
+
 def split_tensor_along_last_dim(
         tensor: torch.Tensor,
         num_partitions: int,
@@ -87,12 +119,12 @@ def split_tensor_along_last_dim(
 class RotaryEmbedding(nn.Module):
     def __init__(self, dim, original_impl=False, device=None, dtype=None):
         super().__init__()
-        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device, dtype=dtype) / dim))
+        inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device).to(dtype=dtype) / dim))
         self.register_buffer("inv_freq", inv_freq)
         self.dim = dim
         self.original_impl = original_impl
 
-    def forward_original_impl(
+    def forward_impl(
             self, seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000
     ):
         """Enhanced Transformer with Rotary Position Embedding.
@@ -118,14 +150,13 @@ class RotaryEmbedding(nn.Module):
         return cache
 
     def forward(self, max_seq_len, offset=0):
-        if self.original_impl:
-            return self.forward_original_impl(
-                max_seq_len, self.dim, dtype=self.inv_freq.dtype, device=self.inv_freq.device
-            )
+        return self.forward_impl(
+            max_seq_len, self.dim, dtype=self.inv_freq.dtype, device=self.inv_freq.device
+        )
 
 
 @torch.jit.script
-def apply_rotary_pos_emb_original(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
+def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
     # x: [sq, b, np, hn]
     sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
     rot_dim = rope_cache.shape[-2] * 2
@@ -151,10 +182,12 @@ class RMSNorm(torch.nn.Module):
         self.weight = torch.nn.Parameter(torch.empty(normalized_shape, device=device, dtype=dtype))
         self.eps = eps
 
-    def forward(self, input: torch.Tensor):
-        norm_x = torch.mean(input * input, dim=-1, keepdim=True)
-        x_normed = input * torch.rsqrt(norm_x + self.eps)
-        return self.weight * x_normed
+    def forward(self, hidden_states: torch.Tensor):
+        input_dtype = hidden_states.dtype
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.eps)
+
+        return (self.weight * hidden_states).to(input_dtype)
 
 
 class CoreAttention(torch.nn.Module):
@@ -311,8 +344,6 @@ class SelfAttention(torch.nn.Module):
                                device=device, **_config_to_kwargs(config)
                                )
 
-        self.interleaved_qkv = config.interleaved_qkv
-
     def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None):
         if self.multi_query_attention:
             num_attention_heads = self.num_multi_query_groups_per_partition
@@ -362,40 +393,25 @@ class SelfAttention(torch.nn.Module):
                 + (self.num_multi_query_groups_per_partition, self.hidden_size_per_attention_head)
             )
         else:
-            if self.interleaved_qkv:
-                new_tensor_shape = mixed_x_layer.size()[:-1] + \
-                                   (self.num_attention_heads_per_partition,
-                                    3 * self.hidden_size_per_attention_head)
-                mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
+            new_tensor_shape = mixed_x_layer.size()[:-1] + \
+                               (self.num_attention_heads_per_partition,
+                                3 * self.hidden_size_per_attention_head)
+            mixed_x_layer = mixed_x_layer.view(*new_tensor_shape)
 
             # [sq, b, np, 3 * hn] --> 3 [sq, b, np, hn]
             (query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3)
 
-            if not self.interleaved_qkv:
-                query_layer = query_layer.view(
-                    query_layer.size()[:-1] + (
-                        self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
-                ).contiguous()
-                key_layer = key_layer.view(
-                    key_layer.size()[:-1] + (
-                        self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
-                ).contiguous()
-                value_layer = value_layer.view(
-                    value_layer.size()[:-1] + (
-                        self.num_attention_heads_per_partition, self.hidden_size_per_attention_head)
-                ).contiguous()
-
         # apply relative positional encoding (rotary embedding)
         if rotary_pos_emb is not None:
-            query_layer = apply_rotary_pos_emb_original(query_layer, rotary_pos_emb)
-            key_layer = apply_rotary_pos_emb_original(key_layer, rotary_pos_emb)
+            query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb)
+            key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb)
 
         # adjust key and value for inference
+        if kv_cache is not None:
+            cache_k, cache_v = kv_cache
+            key_layer = torch.cat((cache_k, key_layer), dim=0)
+            value_layer = torch.cat((cache_v, value_layer), dim=0)
         if use_cache:
-            if kv_cache is not None:
-                cache_k, cache_v = kv_cache
-                key_layer = torch.cat((cache_k, key_layer), dim=0)
-                value_layer = torch.cat((cache_v, value_layer), dim=0)
             kv_cache = (key_layer, value_layer)
         else:
             kv_cache = None
@@ -582,6 +598,8 @@ class GLMTransformer(torch.nn.Module):
             self.final_layernorm = LayerNormFunc(config.hidden_size, eps=config.layernorm_epsilon, device=device,
                                                  dtype=config.torch_dtype)
 
+        self.gradient_checkpointing = False
+
     def _get_layer(self, layer_number):
         return self.layers[layer_number]
 
@@ -593,6 +611,13 @@ class GLMTransformer(torch.nn.Module):
         if not kv_caches:
             kv_caches = [None for _ in range(self.num_layers)]
         presents = () if use_cache else None
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
         all_self_attentions = None
         all_hidden_states = () if output_hidden_states else None
         for index in range(self.num_layers):
@@ -600,14 +625,24 @@ class GLMTransformer(torch.nn.Module):
                 all_hidden_states = all_hidden_states + (hidden_states,)
 
             layer = self._get_layer(index)
-
-            hidden_states, kv_cache = layer(
-                hidden_states,
-                attention_mask,
-                rotary_pos_emb,
-                kv_cache=kv_caches[index],
-                use_cache=use_cache
-            )
+            if self.gradient_checkpointing and self.training:
+                layer_ret = torch.utils.checkpoint.checkpoint(
+                    layer,
+                    hidden_states,
+                    attention_mask,
+                    rotary_pos_emb,
+                    kv_caches[index],
+                    use_cache
+                )
+            else:
+                layer_ret = layer(
+                    hidden_states,
+                    attention_mask,
+                    rotary_pos_emb,
+                    kv_cache=kv_caches[index],
+                    use_cache=use_cache
+                )
+            hidden_states, kv_cache = layer_ret
             if use_cache:
                 presents = presents + (kv_cache,)
 
@@ -661,7 +696,7 @@ class ChatGLMPreTrainedModel(PreTrainedModel):
         return position_ids
 
     def _set_gradient_checkpointing(self, module, value=False):
-        if isinstance(module, ChatGLMModel):
+        if isinstance(module, GLMTransformer):
             module.gradient_checkpointing = value
 
 
@@ -704,6 +739,9 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
         if device is not None:
             init_kwargs["device"] = device
         self.embedding = init_method(Embedding, config, **init_kwargs)
+        self.num_layers = config.num_layers
+        self.multi_query_group_num = config.multi_query_group_num
+        self.kv_channels = config.kv_channels
 
         # Rotary positional embeddings
         self.seq_length = config.seq_length
@@ -711,18 +749,37 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
             config.hidden_size // config.num_attention_heads if config.kv_channels is None else config.kv_channels
         )
 
-        if config.rotary_percent < 1.0:
-            rotary_dim = int(rotary_dim * config.rotary_percent)
-
-        # partial rotary embeddings, which is better than full rotary
-        # Wang and Komatsuzaki et al
-        # https://github.com/kingoflolz/mesh-transformer-jax/
-        self.rotary_pos_emb = RotaryEmbedding(rotary_dim, original_impl=config.original_rope, device=device,
+        self.rotary_pos_emb = RotaryEmbedding(rotary_dim // 2, original_impl=config.original_rope, device=device,
                                               dtype=config.torch_dtype)
         self.encoder = init_method(GLMTransformer, config, **init_kwargs)
         self.output_layer = init_method(nn.Linear, config.hidden_size, config.padded_vocab_size, bias=False,
                                         dtype=config.torch_dtype, **init_kwargs)
-        self.gradient_checkpointing = False
+        self.pre_seq_len = config.pre_seq_len
+        self.prefix_projection = config.prefix_projection
+        if self.pre_seq_len is not None:
+            for param in self.parameters():
+                param.requires_grad = False
+            self.prefix_tokens = torch.arange(self.pre_seq_len).long()
+            self.prefix_encoder = PrefixEncoder(config)
+            self.dropout = torch.nn.Dropout(0.1)
+
+    def get_input_embeddings(self):
+        return self.embedding.word_embeddings
+
+    def get_prompt(self, batch_size, device, dtype=torch.half):
+        prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(device)
+        past_key_values = self.prefix_encoder(prefix_tokens).type(dtype)
+        past_key_values = past_key_values.view(
+            batch_size,
+            self.pre_seq_len,
+            self.num_layers * 2,
+            self.multi_query_group_num,
+            self.kv_channels
+        )
+        # seq_len, b, nh, hidden_size
+        past_key_values = self.dropout(past_key_values)
+        past_key_values = past_key_values.permute([2, 1, 0, 3, 4]).split(2)
+        return past_key_values
 
     def forward(
             self,
@@ -747,8 +804,17 @@ class ChatGLMModel(ChatGLMPreTrainedModel):
         if inputs_embeds is None:
             inputs_embeds = self.embedding(input_ids)
 
-        if full_attention_mask is None and attention_mask is not None and not attention_mask.all():
-            full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
+        if self.pre_seq_len is not None:
+            if past_key_values is None:
+                past_key_values = self.get_prompt(batch_size=batch_size, device=input_ids.device,
+                                                  dtype=inputs_embeds.dtype)
+            if attention_mask is not None:
+                attention_mask = torch.cat([attention_mask.new_ones((batch_size, self.pre_seq_len)),
+                                            attention_mask], dim=-1)
+
+        if full_attention_mask is None:
+            if (attention_mask is not None and not attention_mask.all()) or (past_key_values and seq_length != 1):
+                full_attention_mask = self.get_masks(input_ids, past_key_values, padding_mask=attention_mask)
 
         # Rotary positional embeddings
         rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
@@ -820,6 +886,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
                 [position_ids, new_position_id], dim=-1
             )
 
+        model_kwargs["is_first_forward"] = False
         return model_kwargs
 
     def prepare_inputs_for_generation(
@@ -828,20 +895,21 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
             past_key_values: Optional[torch.Tensor] = None,
             attention_mask: Optional[torch.Tensor] = None,
             position_ids: Optional[torch.Tensor] = None,
-            input_pos: int = None,
+            is_first_forward: bool = True,
             **kwargs
     ) -> dict:
         # only last token for input_ids if past is not None
-        if past_key_values is not None:
-            if position_ids is None:
-                position_ids = self.get_position_ids(input_ids, device=input_ids.device)
+        if position_ids is None:
+            position_ids = self.get_position_ids(input_ids, device=input_ids.device)
+        if not is_first_forward:
             position_ids = position_ids[..., -1:]
             input_ids = input_ids[:, -1:]
         return {
             "input_ids": input_ids,
             "past_key_values": past_key_values,
             "position_ids": position_ids,
-            "attention_mask": attention_mask
+            "attention_mask": attention_mask,
+            "return_last_logit": True
         }
 
     def forward(
@@ -856,6 +924,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
             output_attentions: Optional[bool] = None,
             output_hidden_states: Optional[bool] = None,
             return_dict: Optional[bool] = None,
+            return_last_logit: Optional[bool] = False,
     ):
         use_cache = use_cache if use_cache is not None else self.config.use_cache
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
@@ -872,7 +941,8 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
         )
 
         hidden_states = transformer_outputs[0]
-
+        if return_last_logit:
+            hidden_states = hidden_states[-1:]
         lm_logits = self.transformer.output_layer(hidden_states)
         lm_logits = lm_logits.transpose(0, 1).contiguous()
 
@@ -927,16 +997,25 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
         return response
 
     def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = None):
-        prompt = ""
-        for i, (old_query, response) in enumerate(history):
-            prompt += "[Round {}]\n\n问：{}\n\n答：{}\n\n".format(i + 1, old_query, response)
-        prompt += "[Round {}]\n\n问：{}\n\n答：".format(len(history) + 1, query)
+        prompt = tokenizer.build_prompt(query, history=history)
         inputs = tokenizer([prompt], return_tensors="pt")
         inputs = inputs.to(self.device)
         return inputs
 
-    @torch.no_grad()
-    def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048, num_beams=1,
+    def build_stream_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = None):
+        if history:
+            prompt = "\n\n[Round {}]\n\n问：{}\n\n答：".format(len(history) + 1, query)
+            input_ids = tokenizer.encode(prompt, add_special_tokens=False)
+            input_ids = input_ids[1:]
+            inputs = tokenizer.batch_encode_plus([(input_ids, None)], return_tensors="pt", add_special_tokens=False)
+        else:
+            prompt = "[Round {}]\n\n问：{}\n\n答：".format(len(history) + 1, query)
+            inputs = tokenizer([prompt], return_tensors="pt")
+        inputs = inputs.to(self.device)
+        return inputs
+
+    @torch.inference_mode()
+    def chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 8192, num_beams=1,
              do_sample=True, top_p=0.8, temperature=0.8, logits_processor=None, **kwargs):
         if history is None:
             history = []
@@ -953,9 +1032,10 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
         history = history + [(query, response)]
         return response, history
 
-    @torch.no_grad()
-    def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, max_length: int = 2048,
-                    do_sample=True, top_p=0.7, temperature=0.95, logits_processor=None, **kwargs):
+    @torch.inference_mode()
+    def stream_chat(self, tokenizer, query: str, history: List[Tuple[str, str]] = None, past_key_values=None,
+                    max_length: int = 8192, do_sample=True, top_p=0.8, temperature=0.8, logits_processor=None,
+                    return_past_key_values=False, **kwargs):
         if history is None:
             history = []
         if logits_processor is None:
@@ -963,15 +1043,33 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
         logits_processor.append(InvalidScoreLogitsProcessor())
         gen_kwargs = {"max_length": max_length, "do_sample": do_sample, "top_p": top_p,
                       "temperature": temperature, "logits_processor": logits_processor, **kwargs}
-        inputs = self.build_inputs(tokenizer, query, history=history)
-        for outputs in self.stream_generate(**inputs, **gen_kwargs):
+        if past_key_values is None and not return_past_key_values:
+            inputs = self.build_inputs(tokenizer, query, history=history)
+        else:
+            inputs = self.build_stream_inputs(tokenizer, query, history=history)
+        if past_key_values is not None:
+            past_length = past_key_values[0][0].shape[0]
+            if self.transformer.pre_seq_len is not None:
+                past_length -= self.transformer.pre_seq_len
+            inputs.position_ids += past_length
+            attention_mask = inputs.attention_mask
+            attention_mask = torch.cat((attention_mask.new_ones(1, past_length), attention_mask), dim=1)
+            inputs['attention_mask'] = attention_mask
+        for outputs in self.stream_generate(**inputs, past_key_values=past_key_values,
+                                            return_past_key_values=return_past_key_values, **gen_kwargs):
+            if return_past_key_values:
+                outputs, past_key_values = outputs
             outputs = outputs.tolist()[0][len(inputs["input_ids"][0]):]
             response = tokenizer.decode(outputs)
-            response = self.process_response(response)
-            new_history = history + [(query, response)]
-            yield response, new_history
-
-    @torch.no_grad()
+            if response and response[-1] != "�":
+                response = self.process_response(response)
+                new_history = history + [(query, response)]
+                if return_past_key_values:
+                    yield response, new_history, past_key_values
+                else:
+                    yield response, new_history
+
+    @torch.inference_mode()
     def stream_generate(
             self,
             input_ids,
@@ -979,6 +1077,7 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
             logits_processor: Optional[LogitsProcessorList] = None,
             stopping_criteria: Optional[StoppingCriteriaList] = None,
             prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
+            return_past_key_values=False,
             **kwargs,
     ):
         batch_size, input_ids_seq_length = input_ids.shape[0], input_ids.shape[-1]
@@ -1067,11 +1166,13 @@ class ChatGLMForConditionalGeneration(ChatGLMPreTrainedModel):
                 outputs, model_kwargs, is_encoder_decoder=self.config.is_encoder_decoder
             )
             unfinished_sequences = unfinished_sequences.mul((sum(next_tokens != i for i in eos_token_id)).long())
-
+            if return_past_key_values:
+                yield input_ids, outputs.past_key_values
+            else:
+                yield input_ids
             # stop when each sentence is finished, or if we exceed the maximum length
             if unfinished_sequences.max() == 0 or stopping_criteria(input_ids, scores):
                 break
-            yield input_ids
 
     def quantize(self, bits: int, empty_init=False, device=None, **kwargs):
         if bits == 0:

quantization.py

@@ -24,7 +24,7 @@ try:
             for name in self._function_names:
                 setattr(self, name, KernelFunction(self._cmodule, name))
 
-    quantization_code = "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"
+    quantization_code = "$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"
 
     kernels = Kernel(
         bz2.decompress(base64.b64decode(quantization_code)),
@@ -32,10 +32,8 @@ try:
             "int4WeightCompression",
             "int4WeightExtractionFloat",
             "int4WeightExtractionHalf",
-            "int4WeightExtractionBFloat16",
             "int8WeightExtractionFloat",
             "int8WeightExtractionHalf",
-            "int8WeightExtractionBFloat16",
         ],
     )
 except Exception as exception:

save_model.py

@@ -0,0 +1,4 @@
+from transformers import AutoModel
+
+model = AutoModel.from_pretrained("/mnt/vepfs/qinkai/release/codegeex2-6b/", trust_remote_code=True).cuda()
+model.save_pretrained("./", max_shard_size="2000MB")

tokenization_chatglm.py

@@ -17,7 +17,7 @@ class SPTokenizer:
         self.n_words: int = self.sp_model.vocab_size()
         self.bos_id: int = self.sp_model.bos_id()
         self.eos_id: int = self.sp_model.eos_id()
-        self.pad_id: int = self.sp_model.eos_id()
+        self.pad_id: int = self.sp_model.unk_id()
         assert self.sp_model.vocab_size() == self.sp_model.get_piece_size()
 
         special_tokens = ["[MASK]", "[gMASK]", "[sMASK]", "sop", "eop"]
@@ -55,7 +55,7 @@ class SPTokenizer:
 
     def convert_id_to_token(self, index):
         """Converts an index (integer) in a token (str) using the vocab."""
-        if index in self.index_special_tokens:
+        if index in self.index_special_tokens or index in [self.eos_id, self.bos_id, self.pad_id] or index < 0:
             return ""
         return self.sp_model.IdToPiece(index)
 
@@ -65,10 +65,11 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
 
     model_input_names = ["input_ids", "attention_mask", "position_ids"]
 
-    def __init__(self, vocab_file, padding_side="left", **kwargs):
-        super().__init__(padding_side=padding_side, clean_up_tokenization_spaces=False, **kwargs)
+    def __init__(self, vocab_file, padding_side="left", clean_up_tokenization_spaces=False, **kwargs):
+        super().__init__(padding_side=padding_side, clean_up_tokenization_spaces=clean_up_tokenization_spaces, **kwargs)
         self.name = "GLMTokenizer"
 
+        self.vocab_file = vocab_file
         self.tokenizer = SPTokenizer(vocab_file)
         self.special_tokens = {
             "<bos>": self.tokenizer.bos_id,
@@ -82,14 +83,26 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
         assert token in self.tokenizer.special_tokens, f"{token} is not a special token for {self.name}"
         return self.tokenizer.special_tokens[token]
 
+    @property
+    def unk_token(self) -> str:
+        return "<unk>"
+
     @property
     def pad_token(self) -> str:
-        return "</s>"
+        return "<unk>"
 
     @property
     def pad_token_id(self):
         return self.get_command("<pad>")
 
+    @property
+    def eos_token(self) -> str:
+        return "</s>"
+
+    @property
+    def eos_token_id(self):
+        return self.get_command("<eos>")
+
     @property
     def vocab_size(self):
         return self.tokenizer.n_words
@@ -146,6 +159,15 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
         prefix_tokens = [self.get_command("[gMASK]"), self.get_command("sop")]
         return prefix_tokens
 
+    def build_prompt(self, query, history=None):
+        if history is None:
+            history = []
+        prompt = ""
+        for i, (old_query, response) in enumerate(history):
+            prompt += "[Round {}]\n\n问：{}\n\n答：{}\n\n".format(i + 1, old_query, response)
+        prompt += "[Round {}]\n\n问：{}\n\n答：".format(len(history) + 1, query)
+        return prompt
+
     def build_inputs_with_special_tokens(
             self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
     ) -> List[int]:
@@ -217,12 +239,11 @@ class ChatGLMTokenizer(PreTrainedTokenizer):
         needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length
 
         # Initialize attention mask if not present.
-        if max_length is not None:
-            if "attention_mask" not in encoded_inputs:
-                encoded_inputs["attention_mask"] = [1] * seq_length
+        if "attention_mask" not in encoded_inputs:
+            encoded_inputs["attention_mask"] = [1] * seq_length
 
-            if "position_ids" not in encoded_inputs:
-                encoded_inputs["position_ids"] = list(range(seq_length))
+        if "position_ids" not in encoded_inputs:
+            encoded_inputs["position_ids"] = list(range(seq_length))
 
         if needs_to_be_padded:
             difference = max_length - len(required_input)