Update configuration_sdar.py

configuration_sdar.py

@@ -1,171 +1,205 @@
-# SPDX-License-Identifier: Apache-2.0
-# adapted fromhttps://github.com/Gen-Verse/dLLM-RL
-# adapted from SADR https://github.com/JetAstra/SDAR/blob/main/generate.py
-
-import torch
-from torch.nn import functional as F
-from transformers.cache_utils import DynamicCache
-
-
-def top_k_logits(logits, k):
-    if k <= 0:
-        return logits
-    else:
-        values, _ = torch.topk(logits, k)
-        min_values = values[..., -1, None]
-        return torch.where(logits < min_values, torch.full_like(logits, float("-inf")), logits)
-
-
-def top_p_logits(logits, p):
-    sorted_logits, sorted_indices = torch.sort(logits, descending=True)
-    cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
-    sorted_mask = cumulative_probs > p
-    sorted_mask[..., 1:] = sorted_mask[..., :-1].clone()
-    sorted_mask[..., 0] = False
-    mask_indices = torch.scatter(torch.full_like(logits, False, dtype=torch.bool), -1, sorted_indices, sorted_mask)
-    logits = logits.masked_fill(mask_indices, float("-inf"))
-    return logits
-
-
-def sample_with_temperature_topk_topp(logits, temperature=1.0, top_k=0, top_p=1.0):
-    orig_shape = logits.shape[:-1]  # [batch, block]
-    vocab_size = logits.shape[-1]
-
-    logits = logits.reshape(-1, vocab_size)  # [batch*block, vocab]
-
-    if temperature != 1.0:
-        logits = logits / temperature
-    if top_k > 0:
-        logits = top_k_logits(logits, top_k)
-    if top_p < 1.0:
-        logits = top_p_logits(logits, top_p)
-    probs = F.softmax(logits, dim=-1)  # shape: [batch*block, vocab]
-    assert probs.dim() == 2
-    token = torch.multinomial(probs, num_samples=1)  # [batch*block, 1]
-    token_prob = torch.gather(probs, -1, token)  # [batch*block, 1]
-
-    return token.view(*orig_shape), token_prob.view(*orig_shape)
-
-
-def get_num_transfer_tokens(block_length, steps):
-    base = block_length // steps
-    remainder = block_length % steps
-    num_transfer_tokens = torch.zeros(steps, dtype=torch.int64) + base
-    num_transfer_tokens[:remainder] += 1
-    return num_transfer_tokens
-
-
-@torch.no_grad()
-def block_diffusion_generate(
-    model,
-    prompt,
-    mask_id,
-    gen_length=128,
-    block_length=8,
-    denoising_steps=8,
-    temperature=1.0,
-    top_k=0,
-    top_p=1.0,
-    remasking_strategy="low_confidence_dynamic",
-    confidence_threshold=0.85,
-    stopping_criteria_idx=None,
-):
-    model.eval()
-    input_ids = prompt["input_ids"]
-    prompt_length = input_ids.shape[1]
-    past_key_values = DynamicCache()
-
-    num_blocks = (prompt_length + gen_length + block_length - 1) // block_length
-    total_length = num_blocks * block_length
-
-    block_mask = torch.tril(torch.ones(num_blocks, num_blocks, device=model.device))
-    block_diffusion_attention_mask = block_mask.repeat_interleave(block_length, dim=0).repeat_interleave(block_length, dim=1).unsqueeze(0)
-    position_ids = torch.arange(total_length, device=model.device).unsqueeze(0)
-
-    x = torch.full((1, total_length), mask_id, dtype=torch.long, device=model.device)
-    x[:, :prompt_length] = input_ids
-    prefill_blocks = prompt_length // block_length
-    prefill_length = prefill_blocks * block_length
-
-    # Prefill stage
-    if prefill_length > 0:
-        cur_x = x[:, :prefill_length]
-        cur_attn_mask = block_diffusion_attention_mask[:, :prefill_length, :prefill_length]
-        if cur_attn_mask.dim() == 3:
-            cur_attn_mask = cur_attn_mask[:, None, :, :]
-        cur_position_ids = position_ids[:, :prefill_length]
-        model(cur_x, attention_mask=cur_attn_mask, position_ids=cur_position_ids, past_key_values=past_key_values, use_cache=True, store_kv=True)
-
-    num_transfer_tokens = get_num_transfer_tokens(block_length, denoising_steps)
-
-    # Decode stage
-    for num_block in range(prefill_blocks, num_blocks):
-        cur_x = x[:, num_block * block_length : (num_block + 1) * block_length].clone()
-        cur_attn_mask = block_diffusion_attention_mask[:, num_block * block_length : (num_block + 1) * block_length, : (num_block + 1) * block_length]
-        if cur_attn_mask.dim() == 3:
-            cur_attn_mask = cur_attn_mask[:, None, :, :]
-        cur_position_ids = position_ids[:, num_block * block_length : (num_block + 1) * block_length]
-        for step in range(denoising_steps + 1):
-            mask_index = cur_x == mask_id
-            if mask_index.sum() == 0:
-                # Store kv cache
-                model(cur_x, attention_mask=cur_attn_mask, position_ids=cur_position_ids, past_key_values=past_key_values, use_cache=True, store_kv=True)
-                break
-
-            # Denosing
-            output = model(cur_x, attention_mask=cur_attn_mask, position_ids=cur_position_ids, past_key_values=past_key_values, use_cache=True, store_kv=False)
-            # Extract logits from the output - handle both CausalLMOutputWithPast and BaseModelOutputWithPast
-            if hasattr(output, "logits") and output.logits is not None:
-                logits = output.logits
-            elif hasattr(output, "last_hidden_state"):
-                # If logits don't exist but we have hidden states, compute logits from the model's lm_head
-                # This can happen if the model returns BaseModelOutputWithPast instead of CausalLMOutputWithPast
-                if hasattr(model, "lm_head"):
-                    hidden_states = output.last_hidden_state
-                    logits = model.lm_head(hidden_states)
-                else:
-                    raise ValueError("Model output does not contain logits and model does not have lm_head to compute them.")
-            else:
-                raise ValueError(f"Unexpected model output type: {type(output)}. Expected CausalLMOutputWithPast or BaseModelOutputWithPast with logits or last_hidden_state.")
-
-            # Sampling
-            x0, x0_p = sample_with_temperature_topk_topp(logits, temperature=temperature, top_k=top_k, top_p=top_p)
-
-            # Sampling strategy
-            if remasking_strategy == "sequential":
-                transfer_index = torch.zeros_like(x0, dtype=torch.bool)
-                for j in range(cur_x.shape[0]):
-                    if mask_index[j].any():
-                        first_mask_index = mask_index[j].nonzero(as_tuple=True)[0].min().item()
-                        transfer_index[j, first_mask_index : first_mask_index + num_transfer_tokens[step]] = True
-                    else:
-                        raise ValueError("No mask tokens found in the current block.")
-
-            elif remasking_strategy == "low_confidence_static":
-                confidence = torch.where(mask_index, x0_p, -torch.inf)
-                transfer_index = torch.zeros_like(x0, dtype=torch.bool)
-                for j in range(confidence.shape[0]):
-                    _, idx = torch.topk(confidence[j], num_transfer_tokens[step])
-                    transfer_index[j, idx] = True
-
-            elif remasking_strategy == "low_confidence_dynamic":
-                confidence = torch.where(mask_index, x0_p, -torch.inf)
-                transfer_index = torch.zeros_like(x0, dtype=torch.bool)
-                for j in range(confidence.shape[0]):
-                    high_conf_mask = confidence[j] > confidence_threshold
-                    num_high_confidence = high_conf_mask.sum()
-                    if num_high_confidence >= num_transfer_tokens[step]:
-                        transfer_index[j] = high_conf_mask
-                    else:
-                        _, idx = torch.topk(confidence[j], num_transfer_tokens[step])
-                        transfer_index[j, idx] = True
-            else:
-                raise ValueError(f"Unknown remasking strategy: {remasking_strategy}")
-
-            cur_x[transfer_index] = x0[transfer_index]
-
-        x[:, num_block * block_length : (num_block + 1) * block_length] = cur_x
-        if stopping_criteria_idx is not None and any(stop_idx in x[:, prompt_length:] for stop_idx in stopping_criteria_idx):
-            break
-
-    return x
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""SDAR model configuration"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class SDARConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`SDARModel`]. It is used to instantiate a
+    SDAR model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of
+    SDAR-1.7B [DiffuOpen/SDAR-1.7B-Chat](https://huggingface.co/DiffuOpen/SDAR-1.7B-Chat/).
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+    Args:
+        vocab_size (`int`, *optional*, defaults to 151936):
+            Vocabulary size of the SDAR model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`SDARModel`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 22016):
+            Dimension of the MLP representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_key_value_heads (`int`, *optional*, defaults to 32):
+            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+            by meanpooling all the original heads within that group. For more details checkout [this
+            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+        head_dim (`int`, *optional*, defaults to 128):
+            The attention head dimension.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in the decoder.
+        max_position_embeddings (`int`, *optional*, defaults to 32768):
+            The maximum sequence length that this model might ever be used with.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        rms_norm_eps (`float`, *optional*, defaults to 1e-06):
+            The epsilon used by the rms normalization layers.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether the model's input and output word embeddings should be tied.
+        rope_theta (`float`, *optional*, defaults to 10000.0):
+            The base period of the RoPE embeddings.
+        rope_scaling (`Dict`, *optional*):
+            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+            accordingly.
+            Expected contents:
+                `rope_type` (`str`):
+                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+                    'llama3'], with 'default' being the original RoPE implementation.
+                `factor` (`float`, *optional*):
+                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+                    original maximum pre-trained length.
+                `original_max_position_embeddings` (`int`, *optional*):
+                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+                    pretraining.
+                `attention_factor` (`float`, *optional*):
+                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+                    computation. If unspecified, it defaults to value recommended by the implementation, using the
+                    `factor` field to infer the suggested value.
+                `beta_fast` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 32.
+                `beta_slow` (`float`, *optional*):
+                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+                    ramp function. If unspecified, it defaults to 1.
+                `short_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `long_factor` (`List[float]`, *optional*):
+                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+                    size divided by the number of attention heads divided by 2
+                `low_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+                `high_freq_factor` (`float`, *optional*):
+                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+        attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`):
+            Whether to use a bias in the query, key, value and output projection layers during self-attention.
+        use_sliding_window (`bool`, *optional*, defaults to `False`):
+            Whether to use sliding window attention.
+        sliding_window (`int`, *optional*, defaults to 4096):
+            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+        max_window_layers (`int`, *optional*, defaults to 28):
+            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    ```python
+    >>> from transformers import SDARModel, SDARConfig
+    >>> # Initializing a SDAR style configuration
+    >>> configuration = SDARConfig()
+    >>> # Initializing a model from the SDAR-8B style configuration
+    >>> model = SDARModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "sdar"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    # Default tensor parallel plan for base model `SDAR`
+    base_model_tp_plan = {
+        "layers.*.self_attn.q_proj": "colwise",
+        "layers.*.self_attn.k_proj": "colwise",
+        "layers.*.self_attn.v_proj": "colwise",
+        "layers.*.self_attn.o_proj": "rowwise",
+        "layers.*.mlp.gate_proj": "colwise",
+        "layers.*.mlp.up_proj": "colwise",
+        "layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=151936,
+        hidden_size=4096,
+        intermediate_size=22016,
+        num_hidden_layers=32,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        head_dim=128,
+        hidden_act="silu",
+        max_position_embeddings=32768,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        use_sliding_window=False,
+        sliding_window=4096,
+        max_window_layers=28,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window  # we check `use_sliding_window` in the modeling code
+        self.max_window_layers = max_window_layers
+
+        # for backward compatibility
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.head_dim = head_dim
+        self.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.use_cache = use_cache
+        self.rope_theta = rope_theta
+        self.rope_scaling = rope_scaling
+        self.attention_bias = attention_bias
+        self.attention_dropout = attention_dropout
+        # Validate the correctness of rotary position embeddings parameters
+        # BC: if there is a 'type' field, move it to 'rope_type'.
+        if self.rope_scaling is not None and "type" in self.rope_scaling:
+            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+        rope_config_validation(self)
+
+        super().__init__(
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )
+
+
+__all__ = ["SDARConfig"]