enable DeepSeek-V2

README.md

@@ -240,7 +240,7 @@ The following model architectures, tasks and device distributions have been vali
 | Mllama     | <div style="text-align:left"><li>LoRA</li></div> | :heavy_check_mark: | <li>[image to text](https://github.com/huggingface/optimum-habana/tree/main/examples/image-to-text)</li> |
 | MiniCPM3 |   | <div style="text-align:left"><li>Single card</li></div> | <li>[text generation](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation)</li> |
 | Baichuan2 |   | <div style="text-align:left"><li>Single card</li></div> | <li>[text generation](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation)</li> |
-
+| DeepSeek-V2 |   | :heavy_check_mark: | <li>[text generation](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation)</li> |
 </div>
 
 - Diffusers:

docs/source/index.mdx

@@ -107,6 +107,7 @@ In the tables below, ✅ means single-card, multi-card and DeepSpeed have all be
 | Mllama     | <div style="text-align:left"><li>LoRA</li></div> |✅      | <li>[image to text](https://github.com/huggingface/optimum-habana/tree/main/examples/image-to-text)</li> |
 | MiniCPM3 |   | <div style="text-align:left"><li>Single card</li></div> | <li>[text generation](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation)</li> |
 | Baichuan2 |   | <div style="text-align:left"><li>Single card</li></div> | <li>[text generation](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation)</li> |
+| DeepSeek-V2 |   | ✅ | <li>[text generation](https://github.com/huggingface/optimum-habana/tree/main/examples/text-generation)</li> |
 
 - Diffusers
 

examples/text-generation/run_generation.py

@@ -306,9 +306,9 @@ def setup_parser(parser):
     parser.add_argument(
         "--parallel_strategy",
         type=str,
-        choices=["tp", "none"],  # Add other strategies as needed
+        choices=["tp", "ep", "none"],  # Add other strategies as needed
         default="none",
-        help="Run multi card with the specified parallel strategy. Choices are 'tp' for Tensor Parallel Strategy or 'none'.",
+        help="Run multi card with the specified parallel strategy. Choices are 'tp' for Tensor Parallel Strategy or 'ep' for Expert Parallel Strategy or 'none'.",
     )
     parser.add_argument(
         "--input_embeds",

examples/text-generation/utils.py

@@ -387,6 +387,44 @@ def setup_distributed_model_tp(args, model_dtype, model_kwargs, logger, cache_di
     return model, args.assistant_model
 
 
+def setup_distributed_model_ep(args, model_dtype, model_kwargs, logger):
+    logger.info("Multi-device ep run.")
+
+    assert args.quant_config == "", "Fp8 is not enabled, unset QUANT_CONFIG"
+    assert args.assistant_model is None, "Assistant model must be None"
+
+    from torch import distributed as dist
+
+    if args.device == "hpu":
+        dist.init_process_group(backend="hccl")
+    else:
+        assert False, "Supports EP only on HPU"
+
+    torch._C._distributed_c10d._register_process_group("default", dist.group.WORLD)
+    logger.info("Creating Model")
+    config = AutoConfig.from_pretrained(args.model_name_or_path, torch_dtype=model_dtype, **model_kwargs)
+    config.update({"ep_size": args.world_size})
+
+    model = AutoModelForCausalLM.from_pretrained(
+        args.model_name_or_path,
+        config=config,
+        torch_dtype=model_dtype,
+        **model_kwargs,
+    )
+
+    model = model.eval().to(args.device)
+
+    if args.use_hpu_graphs:
+        from habana_frameworks.torch.hpu import wrap_in_hpu_graph
+
+        model = wrap_in_hpu_graph(model)
+
+    if args.torch_compile:
+        model = get_torch_compiled_model(model)
+
+    return model, args.assistant_model
+
+
 def setup_distributed_model(args, model_dtype, model_kwargs, logger):
     import deepspeed
 
@@ -677,8 +715,10 @@ def initialize_model(args, logger):
         setup_model(args, model_dtype, model_kwargs, logger)
         if not use_deepspeed
         else setup_distributed_model(args, model_dtype, model_kwargs, logger)
-        if not args.parallel_strategy == "tp"
+        if args.parallel_strategy == "none"
         else setup_distributed_model_tp(args, model_dtype, model_kwargs, logger, cache_dir)
+        if args.parallel_strategy == "tp"
+        else setup_distributed_model_ep(args, model_dtype, model_kwargs, logger)
     )
 
     tokenizer, model, assistant_model = setup_tokenizer(args, model, assistant_model, logger)

optimum/habana/transformers/generation/utils.py

@@ -114,6 +114,7 @@
     "mllama",
     "minicpm3",
     "baichuan",
+    "deepseek_v2",
 ]
 
 

optimum/habana/transformers/modeling_utils.py

@@ -33,6 +33,9 @@
     BaichuanTokenizer,
     DeciLMConfig,
     DeciLMForCausalLM,
+    DeepseekTokenizerFast,
+    DeepseekV2Config,
+    DeepseekV2ForCausalLM,
     Gaudi2Idefics2ImageProcessor,
     GaudiBloomForCausalLM,
     GaudiBloomMLP,
@@ -683,6 +686,10 @@ def adapt_transformers_to_gaudi():
     transformers.AutoConfig.register("deci", DeciLMConfig)
     transformers.AutoModelForCausalLM.register(DeciLMConfig, DeciLMForCausalLM)
 
+    transformers.AutoConfig.register("deepseek_v2", DeepseekV2Config)
+    transformers.AutoModelForCausalLM.register(DeepseekV2Config, DeepseekV2ForCausalLM)
+    transformers.AutoTokenizer.register(DeepseekV2Config, fast_tokenizer_class=DeepseekTokenizerFast)
+
     # Optimization for cohere on Gaudi
     transformers.models.cohere.modeling_cohere.CohereDecoderLayer = GaudiCohereDecoderLayer
     transformers.models.cohere.modeling_cohere.CohereForCausalLM = GaudiCohereForCausalLM

optimum/habana/transformers/models/__init__.py

@@ -60,6 +60,11 @@
     DeciLMConfig,
     DeciLMForCausalLM,
 )
+from .deepseek_v2 import (
+    DeepseekTokenizerFast,
+    DeepseekV2Config,
+    DeepseekV2ForCausalLM,
+)
 from .detr import gaudi_DetrConvModel_forward
 from .esm import (
     gaudi_esm_for_protein_folding_forward,

optimum/habana/transformers/models/deepseek_v2/__init__.py

@@ -0,0 +1,3 @@
+from .configuration_deepseek_v2 import DeepseekV2Config
+from .modeling_deepseek_v2 import DeepseekV2ForCausalLM
+from .tokenization_deepseek_v2 import DeepseekTokenizerFast

optimum/habana/transformers/models/deepseek_v2/configuration_deepseek_v2.py

@@ -0,0 +1,206 @@
+"""
+DeepSeekV2 model configuration. Copied from https://huggingface.co/deepseek-ai/DeepSeek-V2-Lite/resolve/main/configuration_deepseek.py"""
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DEEPSEEK_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
+
+
+class DeepseekV2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`DeepseekV2Model`]. It is used to instantiate an DeepSeek
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the DeepSeek-V2.
+    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 102400):
+            Vocabulary size of the Deep model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`DeepseekV2Model`]
+        hidden_size (`int`, *optional*, defaults to 4096):
+            Dimension of the hidden representations.
+        intermediate_size (`int`, *optional*, defaults to 11008):
+            Dimension of the MLP representations.
+        moe_intermediate_size (`int`, *optional*, defaults to 1407):
+            Dimension of the MoE representations.
+        num_hidden_layers (`int`, *optional*, defaults to 32):
+            Number of hidden layers in the Transformer decoder.
+        num_attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        n_shared_experts (`int`, *optional*, defaults to None):
+            Number of shared experts, None means dense model.
+        n_routed_experts (`int`, *optional*, defaults to None):
+            Number of routed experts, None means dense model.
+        routed_scaling_factor (`float`, *optional*, defaults to 1.0):
+            Scaling factor or routed experts.
+        topk_method (`str`, *optional*, defaults to `gready`):
+            Topk method used in routed gate.
+        n_group (`int`, *optional*, defaults to None):
+            Number of groups for routed experts.
+        topk_group (`int`, *optional*, defaults to None):
+            Number of selected groups for each token(for each token, ensuring the selected experts is only within `topk_group` groups).
+        num_experts_per_tok (`int`, *optional*, defaults to None):
+            Number of selected experts, None means dense model.
+        moe_layer_freq (`int`, *optional*, defaults to 1):
+            The frequency of the MoE layer: one expert layer for every `moe_layer_freq - 1` dense layers.
+        first_k_dense_replace (`int`, *optional*, defaults to 0):
+            Number of dense layers in shallow layers(embed->dense->dense->...->dense->moe->moe...->lm_head).
+                                                            \--k dense layers--/
+        norm_topk_prob (`bool`, *optional*, defaults to False):
+            Whether to normalize the weights of the routed experts.
+        scoring_func (`str`, *optional*, defaults to 'softmax'):
+            Method of computing expert weights.
+        aux_loss_alpha (`float`, *optional*, defaults to 0.001):
+            Auxiliary loss weight coefficient.
+        seq_aux = (`bool`, *optional*, defaults to True):
+            Whether to compute the auxiliary loss for each individual sample.
+        num_key_value_heads (`int`, *optional*):
+            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
+            `num_attention_heads`.
+        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 2048):
+            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`.
+        pad_token_id (`int`, *optional*):
+            Padding token id.
+        bos_token_id (`int`, *optional*, defaults to 1):
+            Beginning of stream token id.
+        eos_token_id (`int`, *optional*, defaults to 2):
+            End of stream token id.
+        pretraining_tp (`int`, *optional*, defaults to 1):
+            Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this
+            document](https://huggingface.co/docs/transformers/parallelism) to understand more about it. This value is
+            necessary to ensure exact reproducibility of the pretraining results. Please refer to [this
+            issue](https://github.com/pytorch/pytorch/issues/76232).
+        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+            Whether to tie weight embeddings
+        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. Currently supports two scaling
+            strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
+            `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
+            `max_position_embeddings` to the expected new maximum.
+        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.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+    ```python
+    >>> from transformers import DeepseekV2Model, DeepseekV2Config
+    >>> # Initializing a Deepseek-V2 style configuration
+    >>> configuration = DeepseekV2Config()
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "deepseek_v2"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    def __init__(
+        self,
+        vocab_size=102400,
+        hidden_size=4096,
+        intermediate_size=11008,
+        moe_intermediate_size=1407,
+        num_hidden_layers=30,
+        num_attention_heads=32,
+        num_key_value_heads=32,
+        n_shared_experts=None,
+        n_routed_experts=None,
+        ep_size=1,
+        routed_scaling_factor=1.0,
+        kv_lora_rank=512,
+        q_lora_rank=1536,
+        qk_rope_head_dim=64,
+        v_head_dim=128,
+        qk_nope_head_dim=128,
+        topk_method="gready",
+        n_group=None,
+        topk_group=None,
+        num_experts_per_tok=None,
+        moe_layer_freq=1,
+        first_k_dense_replace=0,
+        norm_topk_prob=False,
+        scoring_func="softmax",
+        aux_loss_alpha=0.001,
+        seq_aux=True,
+        hidden_act="silu",
+        max_position_embeddings=2048,
+        initializer_range=0.02,
+        rms_norm_eps=1e-6,
+        use_cache=True,
+        pad_token_id=None,
+        bos_token_id=100000,
+        eos_token_id=100001,
+        pretraining_tp=1,
+        tie_word_embeddings=False,
+        rope_theta=10000.0,
+        rope_scaling=None,
+        attention_bias=False,
+        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.moe_intermediate_size = moe_intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.n_shared_experts = n_shared_experts
+        self.n_routed_experts = n_routed_experts
+        self.ep_size = ep_size
+        self.routed_scaling_factor = routed_scaling_factor
+        self.kv_lora_rank = kv_lora_rank
+        self.q_lora_rank = q_lora_rank
+        self.qk_rope_head_dim = qk_rope_head_dim
+        self.v_head_dim = v_head_dim
+        self.qk_nope_head_dim = qk_nope_head_dim
+        self.topk_method = topk_method
+        self.n_group = n_group
+        self.topk_group = topk_group
+        self.num_experts_per_tok = num_experts_per_tok
+        self.moe_layer_freq = moe_layer_freq
+        self.first_k_dense_replace = first_k_dense_replace
+        self.norm_topk_prob = norm_topk_prob
+        self.scoring_func = scoring_func
+        self.aux_loss_alpha = aux_loss_alpha
+        self.seq_aux = seq_aux
+        # 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.hidden_act = hidden_act
+        self.initializer_range = initializer_range
+        self.rms_norm_eps = rms_norm_eps
+        self.pretraining_tp = pretraining_tp
+        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
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            tie_word_embeddings=tie_word_embeddings,
+            **kwargs,
+        )