optimum-rbln 0.1.12__py3-none-any.whl → 0.1.15__py3-none-any.whl

optimum/rbln/transformers/models/decoderonly/modeling_decoderonly.py

@@ -20,26 +20,24 @@
 # are the intellectual property of Rebellions Inc. and may not be
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
-import functools
-import glob
-import os
-from abc import ABC
+import inspect
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Dict, List, Optional, Tuple, Union
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union
 
-import rebel  # noqa: F401
-import torch  # noqa: F401
-from safetensors.torch import load_file
+import rebel
+import torch
 from transformers import AutoConfig, AutoModelForCausalLM, PretrainedConfig, PreTrainedModel
 from transformers.modeling_utils import no_init_weights
 from transformers.utils import ModelOutput
 
-from ....modeling_base import RBLNModel
+from ....modeling import RBLNModel
 from ....modeling_config import DEFAULT_COMPILED_MODEL_NAME, RBLNCompileConfig, RBLNConfig
 from ....utils.logging import get_logger
 from ....utils.runtime_utils import RBLNPytorchRuntime
 from ....utils.timer_utils import rbln_timer
+from ...utils.rbln_quantization import QuantizationManager
+from .decoderonly_architecture import DecoderOnlyWrapper
 
 
 logger = get_logger()
@@ -52,12 +50,6 @@ if TYPE_CHECKING:
         PretrainedConfig,
     )
 
-SUPPORTED_QUANTIZATIONS = {
-    "rbln": [
-        "w4a16",
-    ],
-}
-
 
 class RBLNRuntimeModel(RBLNPytorchRuntime):
     mandatory_members = ["main_input_name", "embed_tokens"]
@@ -102,19 +94,30 @@ class RBLNDecoderOnlyOutput(ModelOutput):
     generate_idx: torch.Tensor = None
 
 
-class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
+class RBLNDecoderOnlyModelForCausalLM(RBLNModel):
     """
-    The DecoderOnly Model transformer with a language modeling head (linear layer) on top.
-    This model inherits from [`RBLNMultiModel`]. Check the superclass documentation for the generic methods the library implements for all its models.
-
-    A class to convert and run pre-trained transformers based DecoderOnlyForCausalLM model on RBLN devices.
-    It implements the methods to convert a pre-trained transformers DecoderOnlyForCausalLM model into a RBLN transformer model by:
-    - transferring the checkpoint weights of the original into an optimized RBLN graph,
-    - compiling the resulting graph using the RBLN compiler.
+    A base class for decoder-only transformer models optimized for causal language modeling tasks on RBLN devices.
+    This class serves as the foundation for various decoder-only architectures like GPT, LLaMA, etc.
+
+    The class provides core functionality for:
+    1. Converting pre-trained transformer models to RBLN-optimized format
+    2. Handling the compilation process for RBLN devices
+    3. Managing inference operations for causal language modeling
+
+    This class inherits from RBLNModel and implements specific methods required for
+    decoder-only architectures and causal language modeling tasks.
+
+    Note:
+        - This class is designed to be subclassed by specific model implementations
+          (e.g., RBLNLlamaForCausalLM, RBLNGPT2LMHeadModel)
+        - Subclasses should implement model-specific conversion logic.
+        - The class handles RBLN-specific optimizations automatically during compilation
     """
 
     main_input_name = "input_ids"
     auto_model_class = AutoModelForCausalLM
+    _decoder_wrapper_cls = DecoderOnlyWrapper
+    _use_rotary_emb = True
 
     def __post_init__(self, **kwargs):
         self.batch_size = self.rbln_config.model_cfg["batch_size"]
@@ -173,6 +176,7 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
     def get_quantized_model(
         cls,
         model_id: str,
+        config: Optional[PretrainedConfig] = None,
         use_auth_token: Optional[Union[bool, str]] = None,
         revision: Optional[str] = None,
         force_download: bool = False,
@@ -182,56 +186,47 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
         trust_remote_code: bool = False,
         **kwargs,
     ):
-        from ...utils.rbln_quantization import update_layers_to_quantized
+        from ...utils.rbln_quantization import prepare_model_for_quantization
 
         kwargs = cls.update_kwargs(kwargs)
 
-        config = AutoConfig.from_pretrained(
-            model_id,
-            use_auth_token=use_auth_token,
-            revision=revision,
-            force_download=force_download,
-            cache_dir=cache_dir,
-            trust_remote_code=trust_remote_code,
-            **kwargs,
-        )
+        if config is None:
+            config = AutoConfig.from_pretrained(
+                model_id,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                force_download=force_download,
+                cache_dir=cache_dir,
+                trust_remote_code=trust_remote_code,
+                **kwargs,
+            )
 
         with no_init_weights():
             model = AutoModelForCausalLM.from_config(config)
 
-        update_layers_to_quantized(model)
-
-        n_layer = kwargs.get("num_hidden_layers", None)
-        cls._load_weights_directly_to_model(model, model_id, n_layer)
+        prepare_model_for_quantization(model, model_id, kwargs.get("num_hidden_layers"))
 
         return model
 
-    def _load_weights_directly_to_model(model, model_id, n_layer=None):
+    def __getattr__(self, __name: str) -> Any:
         """
-        Load safetensor file data directly into the model, filtering by layer if n_layer is provided.
+        Special method to delegate attribute access to the original Huggingface LM class.
+        This method is called when an attribute is not found in the current instance's dictionary.
+        It enables transparent access to the original model's attributes and methods while maintaining
+        proper method binding.
+
+        The method implements a delegation pattern that:
+        1. For methods: Creates a wrapper that properly binds 'self' to method calls
+        2. For other attributes: Returns them directly from the original class
         """
 
-        model_params = dict(model.named_parameters(recurse=True))
-        model_buffers = dict(model.named_buffers(recurse=True))
-        safetensor_files = glob.glob(f"{model_id}/*.safetensors")
-
-        target_layers = list(range(n_layer)) if n_layer is not None else None
+        def redirect(func):
+            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
 
-        for safetensor_file in safetensor_files:
-            file_data = load_file(safetensor_file)
-            for key, value in file_data.items():
-                if target_layers is not None:
-                    parts = key.split(".")
-
-                    if len(parts) > 2 and parts[2].isdigit() and (int(parts[2]) not in target_layers):
-                        continue
-
-                if key in model_params:
-                    model_params[key].data.copy_(value)
-                elif key in model_buffers:
-                    model_buffers[key].data.copy_(value)
-
-        return 0
+        val = getattr(self.hf_class, __name, None) or getattr(PreTrainedModel, __name)
+        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
+            return redirect(val)
+        return val
 
     @classmethod
     def get_pytorch_model(cls, *args, **kwargs) -> "PreTrainedModel":
@@ -245,53 +240,17 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
 
         return model
 
-    def validate_quantization_config(quantize_config):
-        if quantize_config is not None:
-            q_format = quantize_config.get("format")
-            q_precision = quantize_config.get("precision")
-
-            if q_format not in SUPPORTED_QUANTIZATIONS:
-                raise ValueError(
-                    f"Invalid quantization format: {q_format}. "
-                    f"Supported formats are: {list(SUPPORTED_QUANTIZATIONS.keys())}"
-                )
-
-            if q_precision not in SUPPORTED_QUANTIZATIONS[q_format]:
-                raise ValueError(
-                    f"Invalid precision: {q_precision} for format: {q_format}. "
-                    f"Supported precisions are: {SUPPORTED_QUANTIZATIONS[q_format]}"
-                )
-
-        return quantize_config
-
     @classmethod
-    def set_quantize_env(cls, quantize_config):
-        RBLN_QUANT_BITS_ENV = "RBLN_QUANT_BITS"
-        quantize_config = cls.validate_quantization_config(quantize_config)
-        if quantize_config is not None:
-            q_precision = quantize_config.get("precision")
-            quant_bits = q_precision.split("w")[1].split("a")[0]
-            os.environ[RBLN_QUANT_BITS_ENV] = quant_bits
-            return RBLN_QUANT_BITS_ENV
-        return None
+    def wrap_model_if_needed(cls, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
+        wrapper_cfg = {"max_seq_len": rbln_config.model_cfg["max_seq_len"]}
 
-    @classmethod
-    def reset_quantize_env(cls, env_var_name):
-        if env_var_name is not None and env_var_name in os.environ:
-            del os.environ[env_var_name]
+        # If the model wrapper supports rbln-custom-flash-attention
+        if "kvcache_partition_len" in inspect.signature(cls._decoder_wrapper_cls.__init__).parameters:
+            wrapper_cfg["kvcache_partition_len"] = rbln_config.model_cfg.get("kvcache_partition_len")
 
-    @classmethod
-    def manage_quantize_env(cls, func):
-        @functools.wraps(func)
-        def wrapper(*args, **kwargs):
-            quantize_config = kwargs.get("quantize_config")
-            quantize_env_var = cls.set_quantize_env(quantize_config)
-            try:
-                return func(*args, **kwargs)
-            finally:
-                cls.reset_quantize_env(quantize_env_var)
-
-        return wrapper
+        wrapper_cfg["use_rotary_emb"] = cls._use_rotary_emb
+
+        return cls._decoder_wrapper_cls(model, **wrapper_cfg).eval()
 
     @classmethod
     @torch.inference_mode()
@@ -305,15 +264,15 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
         @rbln_timer("JIT trace")
         def get_scripted_model():
             # This function is nested to dealloc the example inputs before compilation.
+            # FIXME: 3rd dummy_input(batch_idx) should be fill zero to compile flash_attn.
             prefill_example_inputs = prefill_rbln_compile_config.get_dummy_inputs(fill=0)
-            dec_example_inputs = dec_rbln_compile_config.get_dummy_inputs(fill=4)
-
-            batch_index = 3
-            dec_example_inputs[batch_index].fill_(-1)  # fill batch_position -1 to indicate it is decoder.
+            dec_example_inputs = dec_rbln_compile_config.get_dummy_inputs(fill=0)
 
+            wrapped_model.phase = "prefill"
             prefill_scripted_model = torch.jit.trace(
                 wrapped_model, prefill_example_inputs, check_trace=False, _store_inputs=False
             )
+            wrapped_model.phase = "decode"
             dec_scripted_model = torch.jit.trace(
                 wrapped_model, dec_example_inputs, check_trace=False, _store_inputs=False
             )
@@ -336,6 +295,7 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
         prefill_ir, dec_ir = scripted_model_to_ir()
         # Caching prefill_decoder/decoder I/O
         cache_index_offset = 5
+
         connections = [
             (prefill_ir.outputs[1 + i], prefill_ir.inputs[cache_index_offset + i])
             for i in range(model.config.num_hidden_layers * 2)
@@ -344,7 +304,7 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
         # Extract quantize_config from rbln_config
         quantize_config = rbln_config.model_cfg.get("quantization", None)
 
-        @cls.manage_quantize_env
+        @QuantizationManager.with_quantization_env
         def compile_model(*args, **kwargs):
             # Remove quantize_config from kwargs
             kwargs.pop("quantize_config", None)
@@ -374,10 +334,8 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
     ) -> RBLNConfig:
         rbln_max_seq_len = rbln_kwargs.get("max_seq_len", None)
         rbln_batch_size = rbln_kwargs.get("batch_size", None)
-        rbln_quantization = rbln_kwargs.get("quantization", None)
         rbln_use_inputs_embeds = rbln_kwargs.get("use_inputs_embeds", None)
-
-        rbln_quantization = cls.validate_quantization_config(rbln_quantization)
+        rbln_quantization = QuantizationManager.validate_quantization_config(rbln_kwargs.get("quantization", None))
 
         prefill_chunk_size = 128
         if rbln_max_seq_len is None:
@@ -552,8 +510,6 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
         cache_position: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.LongTensor] = None,
         generate_idx: Optional[torch.Tensor] = None,
-        # from llava_next forward args
-        batch_idx: Optional[int] = None,
         **kwargs,
     ) -> Tuple[torch.FloatTensor]:
         # prefll
@@ -575,7 +531,7 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
                     input_ids=input_tensor if inputs_embeds is None else None,
                     inputs_embeds=input_tensor if inputs_embeds is not None else None,
                     cache_position=cache_position,
-                    batch_idx=b_idx if batch_idx is None else batch_idx,  # Llava-next prefill
+                    batch_idx=b_idx,
                 )
                 logits.append(logit)
             logits = torch.cat(logits, dim=0)
@@ -676,11 +632,24 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
         cache_position: torch.Tensor = None,
     ) -> torch.FloatTensor:
         input_tensors = inputs_embeds if inputs_embeds is not None else input_ids
+        if input_tensors is None:
+            raise ValueError("Either `input_ids` or `inputs_embeds` must be provided.")
 
         batch_size = input_tensors.shape[0]
+        if batch_size != self.batch_size:
+            raise RuntimeError(
+                f"Batch size mismatch: got {batch_size}, expected {self.batch_size} (compiled batch size)."
+            )
+
+        if batch_size != cache_position.shape[0]:
+            raise RuntimeError(f"Cache position size mismatch: got {cache_position.shape[0]}, expected {batch_size}.")
 
         for b_idx in range(batch_size):
             decoding_step = cache_position[b_idx].item()
+            if not (0 <= decoding_step < self.dec_attn_mask.shape[-1]):
+                raise ValueError(
+                    f"Decoding step {decoding_step} out of bounds for attention mask with shape {self.dec_attn_mask.shape}."
+                )
             self.dec_attn_mask[b_idx, :, :, decoding_step] = 1
 
         logits, _ = self.decoder(
@@ -693,31 +662,3 @@ class RBLNDecoderOnlyModelForCausalLM(RBLNModel, ABC):
         )
 
         return logits
-
-    def vllm_forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        inputs_embeds: torch.Tensor = None,
-        cache_position: torch.Tensor = None,
-        batch_idx: Optional[int] = None,
-        **kwargs,
-    ) -> Tuple[torch.FloatTensor]:
-        # prefll
-        if cache_position.shape[-1] > 1:
-            logits = self._forward_prefill(
-                input_ids=input_ids,
-                inputs_embeds=inputs_embeds,
-                cache_position=cache_position,
-                batch_idx=batch_idx,
-            )
-        # decoder
-        else:
-            logits = self._forward_decoder(
-                input_ids=input_ids,
-                inputs_embeds=inputs_embeds,
-                cache_position=cache_position,
-            )
-
-        return RBLNDecoderOnlyOutput(
-            logits=logits,
-        )

optimum/rbln/transformers/models/dpt/modeling_dpt.py

@@ -27,7 +27,7 @@ from typing import TYPE_CHECKING, Any, Dict, Iterable, Optional, Union
 from transformers import AutoModelForDepthEstimation
 from transformers.modeling_outputs import DepthEstimatorOutput
 
-from ....modeling_base import RBLNModel
+from ....modeling import RBLNModel
 from ....modeling_config import RBLNCompileConfig, RBLNConfig
 
 

optimum/rbln/transformers/models/exaone/exaone_architecture.py

@@ -20,53 +20,78 @@
 # are the intellectual property of Rebellions Inc. and may not be
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
+from typing import TYPE_CHECKING
 
+import torch.nn as nn
 
-from ...models.decoderonly import (
+from ....utils import logging
+from ...models.decoderonly.decoderonly_architecture import (
     DecoderOnlyAttention,
-    DecoderOnlyDecoderLayer,
+    DecoderOnlyFlashAttention,
+    DecoderOnlyForCausalLM,
+    DecoderOnlyLayer,
     DecoderOnlyModel,
     DecoderOnlyWrapper,
 )
 
 
+if TYPE_CHECKING:
+    from transformers import PreTrainedModel as ExaoneForCausalLM
+
+logger = logging.get_logger(__name__)
+
+
 class ExaoneForCausalLMWrapper(DecoderOnlyWrapper):
     """A wrapper class for the Exaone model with a language modeling head."""
 
-    def __init__(self, model, max_seq_len):
-        super(DecoderOnlyWrapper, self).__init__()
-        self.config = model.config
-        self.model = self.convert_attribute_name(model.transformer)
-        self.lm_head = model.lm_head
-        self.head_dim = self.config.hidden_size // self.config.num_attention_heads
-        self.max_position_embeddings = (
-            self.config.max_position_embeddings if max_seq_len > self.config.max_position_embeddings else max_seq_len
-        )
-        self.max_seq_len = max_seq_len
-        self.rope_scaling = getattr(self.config, "rope_scaling", None)
-        self.rotary_emb = self._init_rope()
-
-    @staticmethod
-    def convert_attribute_name(model):
-        model.embed_tokens = model.wte
-        model.norm = model.ln_f
-        model.layers = model.h
-
-        for layer in model.layers:
-            layer.input_layernorm = layer.ln_1
-            layer.self_attn = layer.attn.attention
-            layer.post_attention_layernorm = layer.ln_2
-            layer.self_attn.o_proj = layer.self_attn.out_proj
-
-        return model
-
-    def get_forward_dict(self):
-        forward_dict = {}
-        forward_dict.update(
-            {
-                "wrapper": DecoderOnlyModel.forward,
-                "model": DecoderOnlyDecoderLayer.forward,
-                "decoder_layer": DecoderOnlyAttention.forward,
-            }
-        )
-        return forward_dict
+    def convert_to_rbln_causal_lm(self, causal_lm: "ExaoneForCausalLM"):
+        new_layers = []
+        for layer in causal_lm.transformer.h:
+            if self.attn_impl == "eager":
+                new_self_attn = ExaoneAttention(layer.attn.attention)
+            elif self.attn_impl == "flash_attn":
+                new_self_attn = ExaoneFlashAttention(
+                    layer.attn.attention, kvcache_partition_len=self.kvcache_partition_len
+                )
+            else:
+                raise NotImplementedError(f"Unknwon attn : {self.attn_impl}")
+
+            new_layer = ExaoneLayer(layer, new_self_attn)
+            new_layers.append(new_layer)
+        new_model = ExaoneModel(causal_lm.transformer, new_layers)
+        new_causal_lm = DecoderOnlyForCausalLM(causal_lm, new_model)
+        return new_causal_lm
+
+
+class ExaoneModel(DecoderOnlyModel):
+    def get_embedding(self) -> nn.Embedding:
+        return self._original_mod.wte
+
+    def get_last_layernorm(self) -> nn.LayerNorm:
+        return self._original_mod.ln_f
+
+
+class ExaoneLayer(DecoderOnlyLayer):
+    def get_pre_attention_layernorm(self) -> nn.LayerNorm:
+        return self._original_mod.ln_1
+
+    def get_post_attention_layernorm(self) -> nn.LayerNorm:
+        return self._original_mod.ln_2
+
+
+class ExaoneAttention(DecoderOnlyAttention):
+    def __post_init__(self):
+        self.q_proj = self._original_mod.q_proj
+        self.k_proj = self._original_mod.k_proj
+        self.v_proj = self._original_mod.v_proj
+        self.o_proj = self._original_mod.out_proj
+        self.num_key_value_heads = self._original_mod.num_key_value_heads
+
+
+class ExaoneFlashAttention(DecoderOnlyFlashAttention):
+    def __post_init__(self):
+        self.q_proj = self._original_mod.q_proj
+        self.k_proj = self._original_mod.k_proj
+        self.v_proj = self._original_mod.v_proj
+        self.o_proj = self._original_mod.out_proj
+        self.num_key_value_heads = self._original_mod.num_key_value_heads

optimum/rbln/transformers/models/exaone/modeling_exaone.py

@@ -21,21 +21,15 @@
 # copied, modified, or distributed without prior written permission
 # from Rebellions Inc.
 
-import inspect
-import logging
-from typing import TYPE_CHECKING, Any, Callable
 
-from ....modeling_config import RBLNConfig
+from transformers import AutoModelForCausalLM
+
+from ....utils import logging
 from ..decoderonly import RBLNDecoderOnlyModelForCausalLM
 from .exaone_architecture import ExaoneForCausalLMWrapper
-from .hf_hub_cached.modeling_exaone import ExaoneForCausalLM
 
 
-logger = logging.getLogger(__name__)
-if TYPE_CHECKING:
-    from transformers import (
-        PreTrainedModel,
-    )
+logger = logging.get_logger(__name__)
 
 
 class RBLNExaoneForCausalLM(RBLNDecoderOnlyModelForCausalLM):
@@ -52,25 +46,8 @@ class RBLNExaoneForCausalLM(RBLNDecoderOnlyModelForCausalLM):
 
     """
 
-    @classmethod
-    def wrap_model_if_needed(self, model: "PreTrainedModel", rbln_config: "RBLNConfig"):
-        rbln_max_seq_len = rbln_config.model_cfg["max_seq_len"]
-        return ExaoneForCausalLMWrapper(model, rbln_max_seq_len).eval()
-
-    def __getattr__(self, __name: str) -> Any:
-        """This is the key method to implement RBLN-Exaone.
-
-        Returns:
-            Any: Exaone's corresponding method
-        """
-
-        def redirect(func):
-            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
-
-        val = getattr(ExaoneForCausalLM, __name)
-        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
-            return redirect(val)
-        return val
+    _decoder_wrapper_cls = ExaoneForCausalLMWrapper
+    _hf_class = AutoModelForCausalLM
 
     @classmethod
     def from_pretrained(cls, *args, **kwargs):