Remove Optimum-habana

Signed-off-by: yuanwu <yuan.wu@intel.com>

backends/gaudi/server/pyproject.toml

@@ -22,10 +22,9 @@ opentelemetry-instrumentation-grpc = "^0.53b0"
 hf-transfer = "^0.1.9"
 sentencepiece = "^0.2.0"
 peft = "^0.15"
-optimum-habana = "1.17"
+transformers = "^4.52.4"
-transformers = "^4.49"
 numpy = "^1.26"
-accelerate = "^0.33"
+accelerate = "^1.7.0"
 outlines= { version = "^0.0.36", optional = true }
 prometheus-client = "^0.21.1"
 py-cpuinfo = "^9.0.0"

backends/gaudi/server/requirements.txt

@@ -1,4 +1,4 @@
-accelerate==0.33.0 ; python_version >= "3.9" and python_version < "3.13"
+accelerate==1.7.0 ; python_version >= "3.9" and python_version < "3.13"
 annotated-types==0.7.0 ; python_version >= "3.9" and python_version < "3.13"
 attrs==25.3.0 ; python_version >= "3.9" and python_version < "3.13"
 certifi==2025.1.31 ; python_version >= "3.9" and python_version < "3.13"
@@ -46,7 +46,6 @@ opentelemetry-instrumentation==0.53b0 ; python_version >= "3.9" and python_versi
 opentelemetry-proto==1.32.0 ; python_version >= "3.9" and python_version < "3.13"
 opentelemetry-sdk==1.32.0 ; python_version >= "3.9" and python_version < "3.13"
 opentelemetry-semantic-conventions==0.53b0 ; python_version >= "3.9" and python_version < "3.13"
-optimum-habana==1.17.0 ; python_version >= "3.9" and python_version < "3.13"
 optimum==1.24.0 ; python_version >= "3.9" and python_version < "3.13"
 outlines==0.0.36 ; python_version >= "3.9" and python_version < "3.13"
 packaging==24.2 ; python_version >= "3.9" and python_version < "3.13"
@@ -76,7 +75,7 @@ sympy==1.13.1 ; python_version >= "3.9" and python_version < "3.13"
 threadpoolctl==3.6.0 ; python_version >= "3.9" and python_version < "3.13"
 tokenizers==0.21.1 ; python_version >= "3.9" and python_version < "3.13"
 tqdm==4.67.1 ; python_version >= "3.9" and python_version < "3.13"
-transformers==4.49.0 ; python_version >= "3.9" and python_version < "3.13"
+transformers==4.52.4 ; python_version >= "3.9" and python_version < "3.13"
 triton==3.2.0 ; python_version >= "3.9" and python_version < "3.13" and platform_system == "Linux" and platform_machine == "x86_64"
 typer==0.15.2 ; python_version >= "3.9" and python_version < "3.13"
 typing-extensions==4.13.2 ; python_version >= "3.9" and python_version < "3.13"

backends/gaudi/server/text_generation_server/models/__init__.py

@@ -5,7 +5,6 @@ import os
 
 from loguru import logger
 from transformers.configuration_utils import PretrainedConfig
-from transformers.models.auto import modeling_auto
 from huggingface_hub import hf_hub_download, HfApi
 from typing import Optional
 from pathlib import Path
@@ -882,72 +881,6 @@ def get_model(
                 trust_remote_code=trust_remote_code,
             )
 
-    from text_generation_server.models.causal_lm import CausalLM
-    from text_generation_server.models.vlm_causal_lm import VlmCausalLM
-    from text_generation_server.models.custom_modeling.mllama import (
-        MllamaForConditionalGeneration,
-    )
-    from text_generation_server.models.custom_modeling.llava_next import (
-        LlavaNextForConditionalGeneration,
-    )
-    from text_generation_server.models.vlm_causal_lm import (
-        VlmCausalLMBatch,
-    )
-
-    VLM_BATCH_TYPES.add(VlmCausalLMBatch)
-
-    from optimum.habana.transformers.modeling_utils import adapt_transformers_to_gaudi
-
-    adapt_transformers_to_gaudi()
-    if SDP_ON_BF16 == 1:
-        torch._C._set_math_sdp_allow_fp16_bf16_reduction(True)
-    if model_type == "gpt_bigcode":
-        from text_generation_server.models.starcoder import StarCoder
-
-        return StarCoder(model_id=model_id, revision=revision, dtype=dtype)
-    if model_type == "bloom":
-        from text_generation_server.models.bloom import BLOOM
-
-        return BLOOM(
-            model_id=model_id,
-            revision=revision,
-            speculator=speculator,
-            dtype=dtype,
-            trust_remote_code=trust_remote_code,
-        )
-
-    if model_type == "llava_next":
-        return VlmCausalLM(
-            model_class=LlavaNextForConditionalGeneration,
-            model_id=model_id,
-            revision=revision,
-            quantize=None,
-            speculator=speculator,
-            dtype=dtype,
-            trust_remote_code=trust_remote_code,
-        )
-
-    if model_type == "mllama":
-        return VlmCausalLM(
-            model_class=MllamaForConditionalGeneration,
-            model_id=model_id,
-            revision=revision,
-            quantize=None,
-            speculator=speculator,
-            dtype=dtype,
-            trust_remote_code=trust_remote_code,
-        )
-
-    if model_type in modeling_auto.MODEL_FOR_CAUSAL_LM_MAPPING_NAMES:
-        return CausalLM(
-            model_id,
-            revision,
-            quantize=quantize,
-            speculator=speculator,
-            dtype=dtype,
-            trust_remote_code=trust_remote_code,
-        )
-
     raise ValueError(f"Unsupported model type {model_type}")
 
 

backends/gaudi/server/text_generation_server/models/bloom.py

@@ -1,52 +0,0 @@
-# Copyright (C) 2024 Habana Labs, Ltd. an Intel Company.
-
-import torch
-
-from typing import Optional, Type
-
-from transformers import PreTrainedTokenizerBase
-
-from text_generation_server.models import CausalLM
-from text_generation_server.models.causal_lm import CausalLMBatch
-from text_generation_server.pb import generate_pb2
-
-
-class BloomCausalLMBatch(CausalLMBatch):
-    @classmethod
-    def from_pb(
-        cls,
-        pb: generate_pb2.Batch,
-        tokenizer: PreTrainedTokenizerBase,
-        dtype: torch.dtype,
-        device: torch.device,
-    ) -> "CausalLMBatch":
-        batch = super().from_pb(
-            pb=pb,
-            tokenizer=tokenizer,
-            dtype=dtype,
-            device=device,
-        )
-        batch.keys_head_dim_last = False
-        return batch
-
-
-class BLOOM(CausalLM):
-    def __init__(
-        self,
-        model_id: str,
-        revision: Optional[str] = None,
-        speculator: Optional[str] = None,
-        dtype: Optional[torch.dtype] = None,
-        trust_remote_code: bool = False,
-    ):
-        super(BLOOM, self).__init__(
-            model_id=model_id,
-            revision=revision,
-            speculator=speculator,
-            dtype=dtype,
-            trust_remote_code=trust_remote_code,
-        )
-
-    @property
-    def batch_type(self) -> Type[CausalLMBatch]:
-        return BloomCausalLMBatch

backends/gaudi/server/text_generation_server/models/custom_modeling/qwen2_5_vl.py

@@ -49,7 +49,8 @@ from text_generation_server.models.custom_modeling.flash_qwen2_modeling import (
 # Copied from: https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_5_vl/processing_qwen2_5_vl.py
 from typing import Union
 from transformers.feature_extraction_utils import BatchFeature
-from transformers.image_utils import ImageInput, VideoInput
+from transformers.image_utils import ImageInput
+from transformers.video_utils import VideoInput
 from transformers.processing_utils import (
     ProcessingKwargs,
     ProcessorMixin,

backends/gaudi/server/text_generation_server/models/galactica.py

@@ -1,156 +0,0 @@
-import re
-import torch
-import torch.distributed
-
-
-from transformers import (
-    PreTrainedTokenizerBase,
-)
-from text_generation_server.models.causal_lm import CausalLMBatch
-from text_generation_server.pb import generate_pb2
-from text_generation_server.utils import (
-    NextTokenChooser,
-    StoppingCriteria,
-)
-from text_generation_server.utils.chunks import concat_text_chunks
-
-# CREDIT: Papers with code => https://github.com/paperswithcode/galai/blob/main/galai/utils.py
-
-# we split individual characters inside special tokens like [START_DNA]
-CUSTOM_SEQ_RE = re.compile(r"(\[START_(DNA|SMILES|I_SMILES|AMINO)])(.*?)(\[END_\2])")
-
-# token added to implement a custom sequence tokenization. This token is added at
-# corpus cleaning step and removed in pretokenization. The digits are added to increase the chance
-# that they do not occur in the corpus. The digits are escaped so that the token does not appear
-# literally in the source code in case we ever include it in the training data.
-SPLIT_MARKER = f"SPL{1}T-TH{1}S-Pl3A5E"
-
-
-def _insert_split_marker(m: re.Match):
-    """
-    Applies split marker based on a regex match of special tokens such as
-    [START_DNA].
-    Parameters
-    ----------
-    n : str
-        Input text to split
-    Returns
-    ----------
-    str - the text with the split token added
-    """
-    start_token, _, sequence, end_token = m.groups()
-    sequence = re.sub(r"(.)", rf"{SPLIT_MARKER}\1", sequence, flags=re.DOTALL)
-    return f"{start_token}{sequence}{SPLIT_MARKER}{end_token}"
-
-
-def escape_custom_split_sequence(text):
-    """
-    Applies custom splitting to the text for GALILEO's tokenization
-    Parameters
-    ----------
-    text : str
-        Input text to split
-    Returns
-    ----------
-    str - the text with the split token added
-    """
-    return CUSTOM_SEQ_RE.sub(_insert_split_marker, text)
-
-
-# END CREDIT
-
-
-class GalacticaCausalLMBatch(CausalLMBatch):
-    @classmethod
-    def from_pb(
-        cls,
-        pb: generate_pb2.Batch,
-        tokenizer: PreTrainedTokenizerBase,
-        dtype: torch.dtype,
-        device: torch.device,
-    ) -> "GalacticaCausalLMBatch":
-        inputs = []
-        next_token_choosers = []
-        stopping_criterias = []
-        prefix_offsets = []
-        top_n_tokens = []
-        read_offsets = []
-        requests_idx_mapping = {}
-
-        # Parse batch
-        max_truncation = 0
-        padding_right_offset = 0
-        max_decode_tokens = 0
-        for i, r in enumerate(pb.requests):
-            requests_idx_mapping[r.id] = i
-            # Add escape_custom_split_sequence to the CausalLMBatch logic
-            inputs.append(
-                escape_custom_split_sequence(concat_text_chunks(r.input_chunks.chunks))
-            )
-            next_token_choosers.append(
-                NextTokenChooser.from_pb(r.parameters, device, tokenizer)
-            )
-            stopping_criteria = StoppingCriteria.from_pb(
-                r.stopping_parameters, tokenizer
-            )
-            stopping_criterias.append(stopping_criteria)
-            top_n_tokens.append(r.top_n_tokens)
-            max_truncation = max(max_truncation, r.truncate)
-            max_decode_tokens += stopping_criteria.max_new_tokens
-            padding_right_offset = max(
-                padding_right_offset, stopping_criteria.max_new_tokens
-            )
-
-        tokenized_inputs = tokenizer(
-            inputs,
-            return_tensors="pt",
-            padding=True,
-            return_token_type_ids=False,
-            truncation=True,
-            max_length=max_truncation,
-        ).to(device)
-        for _ in pb.requests:
-            input_len = tokenized_inputs["input_ids"].shape[1]
-            prefix_offsets.append(0)
-            read_offsets.append(input_len)
-
-        input_lengths = tokenized_inputs["attention_mask"].sum(1)
-        max_input_length = input_lengths.max()
-
-        input_ids = tokenized_inputs["input_ids"]
-        # Allocate maximum attention_mask
-        attention_mask = input_ids.new_zeros(
-            (pb.size, max_input_length + padding_right_offset)
-        )
-        # Copy tokenizer attention_mask into fully allocated attention_mask
-        attention_mask[:, :max_input_length] = tokenized_inputs["attention_mask"]
-
-        position_ids = tokenized_inputs["attention_mask"].long().cumsum(-1) - 1
-        position_ids.masked_fill_(tokenized_inputs["attention_mask"] == 0, 1)
-        all_input_ids = tokenized_inputs["input_ids"].T.split(1, dim=1)
-        top_n_tokens_tensor = torch.tensor(
-            top_n_tokens, device=device, dtype=torch.int64
-        )
-
-        max_tokens = len(inputs) * max_input_length + max_decode_tokens
-
-        return cls(
-            batch_id=pb.id,
-            requests=pb.requests,
-            requests_idx_mapping=requests_idx_mapping,
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            past_key_values=None,
-            all_input_ids=list(all_input_ids),
-            input_lengths=input_lengths.tolist(),
-            prefix_offsets=prefix_offsets,
-            read_offsets=read_offsets,
-            next_token_choosers=next_token_choosers,
-            stopping_criterias=stopping_criterias,
-            top_n_tokens=top_n_tokens,
-            top_n_tokens_tensor=top_n_tokens_tensor,
-            max_input_length=max_input_length.item(),
-            padding_right_offset=padding_right_offset,
-            max_tokens=max_tokens,
-        )

backends/gaudi/server/text_generation_server/models/idefics_causal_lm.py

@@ -1,882 +0,0 @@
-from io import BytesIO
-from PIL import Image
-import torch
-import time
-
-from dataclasses import dataclass
-from opentelemetry import trace
-from transformers import (
-    AutoConfig,
-    AutoProcessor,
-    AutoTokenizer,
-    PreTrainedTokenizerBase,
-    ProcessorMixin,
-)
-from typing import Optional, Tuple, List, Type, Dict
-
-from text_generation_server.models import Model
-from text_generation_server.models.types import (
-    Batch,
-    Tokens,
-    Generation,
-    GeneratedText,
-)
-from text_generation_server.pb import generate_pb2
-from text_generation_server.utils import NextTokenChooser, StoppingCriteria, Sampling
-import torch.distributed
-from text_generation_server.models.custom_modeling.idefics_modeling import (
-    IdeficsForVisionText2Text,
-)
-from text_generation_server.utils import (
-    initialize_torch_distributed,
-    weight_files,
-    Weights,
-)
-from text_generation_server.utils.quantization import get_loader
-
-tracer = trace.get_tracer(__name__)
-
-
-@dataclass
-class IdeficsCausalLMBatch(Batch):
-    batch_id: int
-    requests: List[generate_pb2.Request]
-    requests_idx_mapping: Dict[int, int]
-
-    # Decoder values
-    input_ids: torch.Tensor
-    attention_mask: torch.Tensor
-    position_ids: torch.Tensor
-    pixel_values: Optional[torch.Tensor]
-    image_hidden_states: Optional[torch.Tensor]
-    image_attention_mask: Optional[torch.Tensor]
-    past_key_values: Optional[List[Tuple]]
-
-    # All tokens
-    all_input_ids: List[torch.Tensor]
-
-    # Lengths of all generations present in the batch
-    input_lengths: List[int]
-    prefix_offsets: List[int]
-    read_offsets: List[int]
-
-    # Generation helpers
-    next_token_choosers: List[NextTokenChooser]
-    stopping_criterias: List[StoppingCriteria]
-
-    # Metadata used for padding
-    max_input_length: int
-    padding_right_offset: int
-
-    # Maximum number of tokens this batch will grow to
-    max_tokens: int
-
-    # Past metadata
-    keys_head_dim_last: bool = True
-
-    def to_pb(self) -> generate_pb2.CachedBatch:
-        return generate_pb2.CachedBatch(
-            id=self.batch_id,
-            request_ids=[r.id for r in self.requests],
-            size=len(self),
-            max_tokens=self.max_tokens,
-        )
-
-    @classmethod
-    def from_pb(
-        cls,
-        pb: generate_pb2.Batch,
-        tokenizer: PreTrainedTokenizerBase,
-        dtype: torch.dtype,
-        device: torch.device,
-    ) -> "IdeficsCausalLMBatch":
-        raise NotImplementedError
-
-    @classmethod
-    def from_pb_processor(
-        cls,
-        pb: generate_pb2.Batch,
-        tokenizer: PreTrainedTokenizerBase,
-        processor: ProcessorMixin,  # Hack
-        config,
-        dtype: torch.dtype,
-        device: torch.device,
-    ) -> "IdeficsCausalLMBatch":
-        inputs = []
-        next_token_choosers = []
-        stopping_criterias = []
-        prefix_offsets = []
-        read_offsets = []
-        requests_idx_mapping = {}
-
-        # Parse batch
-        max_truncation = 0
-        padding_right_offset = 0
-        max_decode_tokens = 0
-        for i, r in enumerate(pb.requests):
-            requests_idx_mapping[r.id] = i
-            inputs.append(r.input_chunks.chunks)
-            next_token_choosers.append(
-                NextTokenChooser.from_pb(r.parameters, device, tokenizer)
-            )
-            stopping_criteria = StoppingCriteria.from_pb(
-                r.stopping_parameters, tokenizer
-            )
-            stopping_criterias.append(stopping_criteria)
-            max_truncation = max(max_truncation, r.truncate)
-            max_decode_tokens += stopping_criteria.max_new_tokens
-            padding_right_offset = max(
-                padding_right_offset, stopping_criteria.max_new_tokens
-            )
-
-        # TODO Check impact on idefics
-        prompts = []
-        for inp in inputs:
-            # Each input is encoded into a list, where each element of this input list is either a string or a URL
-            prompt = []
-            for chunk in inp:
-                chunk_type = chunk.WhichOneof("chunk")
-                if chunk_type == "text":
-                    prompt.append(chunk.text)
-                elif chunk_type == "image":
-                    image = Image.open(BytesIO(chunk.image.data))
-                    prompt.append(image)
-                else:
-                    raise RuntimeError(f"Invalid chunk type {chunk_type}")
-            prompts.append(prompt)
-
-        # The processor replaces the call to tokenizer, and
-        # a/ takes care of fetching images from the URL
-        # b/ generate the correct input_ids, attention_mask, pixel_values, image_attention_mask to feed to the model
-        tokenized_inputs = processor(
-            prompts,
-            return_tensors="pt",
-            padding=True,
-            truncation=True,
-            max_length=max_truncation,
-            # TODO Check impact on idefics
-            # add_end_of_utterance_token=False,  # Already taken care of inside the prompts, so bypassing the processor's handling of this token
-        ).to(device)
-        for _ in pb.requests:
-            input_len = tokenized_inputs["input_ids"].shape[1]
-            prefix_offsets.append(
-                input_len - 5
-            )  # To decode without potential fallbacks errors
-            read_offsets.append(
-                input_len
-            )  # To decode without potential fallbacks errors
-
-        input_lengths = tokenized_inputs["attention_mask"].sum(1)
-        max_input_length = input_lengths.max()
-
-        input_ids = tokenized_inputs["input_ids"]
-        pixel_values = tokenized_inputs.get("pixel_values", None)
-        image_hidden_states = None
-        # Allocate maximum attention_mask
-        attention_mask = input_ids.new_zeros(
-            (pb.size, max_input_length + padding_right_offset)
-        )
-        # Copy tokenizer attention_mask into fully allocated attention_mask
-        attention_mask[:, :max_input_length] = tokenized_inputs["attention_mask"]
-        # Do the same for image_attention_mask
-        if pixel_values is None:
-            image_attention_mask = None
-        else:
-            image_attention_mask = input_ids.new_zeros(
-                (
-                    pb.size,
-                    max_input_length + padding_right_offset,
-                    pixel_values.size(1),
-                )
-            )
-            image_attention_mask[:, :max_input_length, :] = tokenized_inputs[
-                "image_attention_mask"
-            ]
-
-        position_ids = tokenized_inputs["attention_mask"].long().cumsum(-1) - 1
-        position_ids.masked_fill_(tokenized_inputs["attention_mask"] == 0, 1)
-        all_input_ids = tokenized_inputs["input_ids"].T.split(
-            1, dim=1
-        )  # It's input_ids but splitted into a tuple of tensors where each tensor is (seq_len, 1) size. It is then transformed into a list
-
-        max_tokens = len(inputs) * (max_input_length + max_decode_tokens)
-
-        return cls(
-            batch_id=pb.id,
-            requests=pb.requests,
-            requests_idx_mapping=requests_idx_mapping,
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            pixel_values=pixel_values,
-            image_hidden_states=image_hidden_states,
-            image_attention_mask=image_attention_mask,
-            past_key_values=None,
-            all_input_ids=list(all_input_ids),
-            input_lengths=input_lengths.tolist(),
-            prefix_offsets=prefix_offsets,
-            read_offsets=read_offsets,
-            next_token_choosers=next_token_choosers,
-            stopping_criterias=stopping_criterias,
-            max_input_length=max_input_length.item(),
-            padding_right_offset=padding_right_offset,
-            max_tokens=max_tokens,
-        )
-
-    @tracer.start_as_current_span("filter")
-    def filter(self, request_ids: List[int]) -> Optional["IdeficsCausalLMBatch"]:
-        # It deletes requests from the batch. For instance when client lost connection
-        if len(request_ids) == 0:
-            raise ValueError("Batch must have at least one request")
-        if len(request_ids) == len(self):
-            return self
-
-        keep_indices = []
-
-        # New values after filtering
-        requests_idx_mapping = {}
-        requests = []
-        input_lengths = []
-        prefix_offsets = []
-        read_offsets = []
-        all_input_ids = []
-        max_input_length = 0
-
-        next_token_choosers = []
-        stopping_criterias = []
-
-        total_remaining_decode_tokens = 0
-        new_padding_right_offset = 0
-
-        for i, request_id in enumerate(request_ids):
-            idx = self.requests_idx_mapping[request_id]
-            requests_idx_mapping[request_id] = i
-            keep_indices.append(idx)
-
-            requests.append(self.requests[idx])
-            prefix_offsets.append(self.prefix_offsets[idx])
-            read_offsets.append(self.read_offsets[idx])
-            all_input_ids.append(self.all_input_ids[idx])
-
-            request_input_length = self.input_lengths[idx]
-            input_lengths.append(request_input_length)
-            max_input_length = max(max_input_length, request_input_length)
-
-            next_token_choosers.append(self.next_token_choosers[idx])
-            stopping_criteria = self.stopping_criterias[idx]
-            stopping_criterias.append(stopping_criteria)
-            remaining_decode_tokens = (
-                stopping_criteria.max_new_tokens - stopping_criteria.current_tokens
-            )
-            total_remaining_decode_tokens += remaining_decode_tokens
-            new_padding_right_offset = max(
-                new_padding_right_offset, remaining_decode_tokens
-            )
-
-        # Apply indices to input_ids, attention mask, past key values and other items that need to be cached
-        input_ids = self.input_ids[keep_indices]
-        position_ids = self.position_ids[keep_indices]
-        self.attention_mask = self.attention_mask[
-            keep_indices,
-            -(self.padding_right_offset + max_input_length) : (
-                self.attention_mask.shape[1] - self.padding_right_offset
-            )
-            + new_padding_right_offset,
-        ]
-        # Do the same for pixel_values and image_attention_mask
-        pixel_values = self.pixel_values[keep_indices]
-        self.image_attention_mask = self.image_attention_mask[
-            keep_indices,
-            -(self.padding_right_offset + max_input_length) : (
-                self.image_attention_mask.shape[1] - self.padding_right_offset
-            )
-            + new_padding_right_offset,
-            :,
-        ]
-        if self.image_hidden_states is None:
-            image_hidden_states = None
-        else:
-            image_hidden_states = self.image_hidden_states[keep_indices]
-
-        # Ensure that past_key_values tensors can be updated in-place
-        if type(self.past_key_values[0]) is tuple:
-            self.past_key_values = [list(layer) for layer in self.past_key_values]
-
-        # Update tensors in-place to allow incremental garbage collection
-        past_kv_length = max_input_length - 1
-        for layer in self.past_key_values:
-            past_keys, past_values = layer
-            if len(past_keys.shape) == 3:
-                # Force past to be of dim [self_size, num_heads, ...] for easy indexing
-                past_keys = past_keys.view(len(self), -1, *past_keys.shape[-2:])
-                past_values = past_values.view(len(self), -1, *past_values.shape[-2:])
-            if self.keys_head_dim_last:
-                layer[0] = past_keys[keep_indices, :, -past_kv_length:, :]
-            else:
-                layer[0] = past_keys[keep_indices, :, :, -past_kv_length:]
-            del past_keys
-            layer[1] = past_values[keep_indices, :, -past_kv_length:, :]
-            del past_values
-
-        max_tokens = len(request_ids) * max_input_length + total_remaining_decode_tokens
-
-        self.requests = requests
-        self.requests_idx_mapping = requests_idx_mapping
-        self.input_ids = input_ids
-        self.pixel_values = pixel_values
-        self.image_hidden_states = image_hidden_states
-        self.position_ids = position_ids
-        self.all_input_ids = all_input_ids
-        self.input_lengths = input_lengths
-        self.prefix_offsets = prefix_offsets
-        self.read_offsets = read_offsets
-        self.next_token_choosers = next_token_choosers
-        self.stopping_criterias = stopping_criterias
-        self.max_input_length = max_input_length
-        self.padding_right_offset = new_padding_right_offset
-        self.max_tokens = max_tokens
-
-        return self
-
-    @classmethod
-    @tracer.start_as_current_span("concatenate")
-    def concatenate(
-        cls, batches: List["IdeficsCausalLMBatch"]
-    ) -> "IdeficsCausalLMBatch":
-        # It adds new requests to the batch
-        # Used for padding
-        total_batch_size = 0
-        max_input_length = 0
-        max_num_images = 0
-        padding_right_offset = 0
-        for batch in batches:
-            total_batch_size += len(batch)
-            max_input_length = max(max_input_length, batch.max_input_length)
-            max_num_images = max(max_num_images, batch.pixel_values.size(1))
-            padding_right_offset = max(padding_right_offset, batch.padding_right_offset)
-
-        # Batch attributes
-        requests = []
-        requests_idx_mapping = {}
-        input_lengths = []
-        prefix_offsets = []
-        read_offsets = []
-        all_input_ids = []
-        next_token_choosers = []
-        stopping_criterias = []
-        max_tokens = 0
-
-        # Batch tensors
-        input_ids = None
-        attention_mask = None
-        position_ids = None
-        pixel_values = None
-        image_hidden_states = None
-        image_attention_mask = None
-        past_key_values = []
-
-        # Used for slicing correctly inside the tensors
-        # Equivalent to a cumsum on batch sizes
-        start_index = 0
-        for i, batch in enumerate(batches):
-            requests.extend(batch.requests)
-            input_lengths.extend(batch.input_lengths)
-            prefix_offsets.extend(batch.prefix_offsets)
-            read_offsets.extend(batch.read_offsets)
-            all_input_ids.extend(batch.all_input_ids)
-            next_token_choosers.extend(batch.next_token_choosers)
-            stopping_criterias.extend(batch.stopping_criterias)
-
-            if i == 0:
-                requests_idx_mapping = batch.requests_idx_mapping
-            else:
-                # We need to offset the mapping for each batch by the cumulative batch size
-                for k, v in batch.requests_idx_mapping.items():
-                    requests_idx_mapping[k] = v + start_index
-
-            # Slicing end index for this batch
-            end_index = start_index + len(batch)
-
-            # We only concatenate batches that did at least one step
-            if batch.past_key_values is None:
-                raise ValueError("only concatenate prefilled batches")
-
-            # Create empty tensor
-            # input_ids is always of shape [batch_size, 1]
-            # We do not need to pad it
-            if input_ids is None:
-                input_ids = batch.input_ids.new_empty((total_batch_size, 1))
-            # Copy to correct indices
-            input_ids[start_index:end_index] = batch.input_ids
-
-            # Create padded tensor
-            if attention_mask is None:
-                attention_mask = batch.attention_mask.new_zeros(
-                    (total_batch_size, max_input_length + padding_right_offset),
-                )
-
-            curr_batch_max_num_images = batch.pixel_values.size(1)
-            if pixel_values is None:
-                pixel_values = batch.pixel_values.new_zeros(
-                    (total_batch_size, max_num_images, 3, 224, 224)
-                )
-            pixel_values[start_index:end_index, :curr_batch_max_num_images] = (
-                batch.pixel_values
-            )
-
-            if image_attention_mask is None:
-                image_attention_mask = batch.image_attention_mask.new_zeros(
-                    (
-                        total_batch_size,
-                        max_input_length + padding_right_offset,
-                        max_num_images,
-                    )
-                )
-
-            # We need to slice the attention mask to remove padding from previous steps
-            # and to remove unused allocated space
-            left_offset = max_input_length - batch.max_input_length
-            batch_left_offset = (
-                batch.attention_mask.shape[1]
-                - batch.max_input_length
-                - batch.padding_right_offset
-            )
-            attention_mask[
-                start_index:end_index,
-                left_offset:-padding_right_offset,
-            ] = batch.attention_mask[
-                :,
-                batch_left_offset : -batch.padding_right_offset,
-            ]
-            image_attention_mask[
-                start_index:end_index,
-                left_offset:-padding_right_offset,
-                :curr_batch_max_num_images,
-            ] = batch.image_attention_mask[
-                :, batch_left_offset : -batch.padding_right_offset, :
-            ]
-
-            # Create empty tensor
-            # position_ids is always of shape [batch_size, 1]
-            if position_ids is None:
-                position_ids = batch.position_ids.new_empty((total_batch_size, 1))
-            position_ids[start_index:end_index] = batch.position_ids
-
-            # Shenanigans to get dimensions because BLOOM outputs a past with a different shape
-            # BLOOM Keys:   [batch_size * num_heads, head_dim, seq_length]
-            # BLOOM Values: [batch_size * num_heads, seq_length, head_dim]
-            # And ensure that we can update tensors in-place
-            if isinstance(batch.past_key_values[0], tuple):
-                batch.past_key_values = [
-                    [t.view(len(batch), -1, *t.shape[-2:]) for t in layer]
-                    for layer in batch.past_key_values
-                ]
-            elif len(batch.past_key_values[0][0].shape) == 3:
-                for layer in batch.past_key_values:
-                    for k, t in enumerate(layer):
-                        layer[k] = t.view(len(batch), -1, *t.shape[-2:])
-
-            # Add eventual padding tokens that were added while concatenating
-            max_tokens += batch.max_tokens + (
-                max_input_length - batch.max_input_length
-            ) * len(batch)
-
-            start_index = end_index
-
-        first_past_kvs = batches[0].past_key_values
-        _, num_heads, padded_sequence_length, head_dim = first_past_kvs[0][1].shape
-
-        padded_past_values_shape = (
-            total_batch_size,
-            num_heads,
-            max_input_length - 1,
-            head_dim,
-        )
-
-        if batches[0].keys_head_dim_last:
-            padded_past_keys_shape = padded_past_values_shape
-        else:
-            # seq_length is last for BLOOM
-            padded_past_keys_shape = (
-                total_batch_size,
-                num_heads,
-                head_dim,
-                max_input_length - 1,
-            )
-
-        # Iterate over attention layers
-        # Concatenate past key values layer by layer to allow incremental garbage collection
-        for j in range(len(first_past_kvs)):
-            padded_past_keys = first_past_kvs[j][0].new_zeros(padded_past_keys_shape)
-            start_index = 0
-            for batch in batches:
-                past_keys = batch.past_key_values[j][0]
-                # Clear reference to the original tensor
-                batch.past_key_values[j][0] = None
-
-                # Slicing end index for this batch
-                end_index = start_index + len(batch)
-                # We slice the keys to remove the padding from previous batches
-                past_seq_len = batch.max_input_length - 1
-                if batch.keys_head_dim_last:
-                    padded_past_keys[start_index:end_index, :, -past_seq_len:, :] = (
-                        past_keys[:, :, -past_seq_len:, :]
-                    )
-                else:
-                    # BLOOM case
-                    padded_past_keys[start_index:end_index, :, :, -past_seq_len:] = (
-                        past_keys[:, :, :, -past_seq_len:]
-                    )
-                del past_keys
-
-                start_index = end_index
-
-            padded_past_values = first_past_kvs[j][1].new_zeros(
-                padded_past_values_shape
-            )
-            start_index = 0
-            for batch in batches:
-                past_values = batch.past_key_values[j][1]
-                # Clear reference to the original tensor
-                batch.past_key_values[j][1] = None
-
-                # Slicing end index for this batch
-                end_index = start_index + len(batch)
-                # We slice the past values to remove the padding from previous batches
-                past_seq_len = batch.max_input_length - 1
-                padded_past_values[start_index:end_index, :, -past_seq_len:, :] = (
-                    past_values[:, :, -past_seq_len:, :]
-                )
-                del past_values
-
-                # Update values
-                start_index = end_index
-
-            past_key_values.append([padded_past_keys, padded_past_values])
-
-        return cls(
-            batch_id=batches[0].batch_id,
-            requests=requests,
-            requests_idx_mapping=requests_idx_mapping,
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-            position_ids=position_ids,
-            pixel_values=pixel_values,
-            image_hidden_states=image_hidden_states,
-            image_attention_mask=image_attention_mask,
-            past_key_values=past_key_values,
-            all_input_ids=all_input_ids,
-            input_lengths=input_lengths,
-            prefix_offsets=prefix_offsets,
-            read_offsets=read_offsets,
-            next_token_choosers=next_token_choosers,
-            stopping_criterias=stopping_criterias,
-            max_input_length=max_input_length,
-            padding_right_offset=padding_right_offset,
-            keys_head_dim_last=batches[0].keys_head_dim_last,
-            max_tokens=max_tokens,
-        )
-
-    def __len__(self):
-        return len(self.requests)
-
-
-class IdeficsCausalLM(Model):
-    def __init__(
-        self,
-        model_id: str,
-        revision: Optional[str] = None,
-        quantize: Optional[str] = None,
-        speculator: Optional[str] = None,
-        dtype: Optional[torch.dtype] = None,
-        trust_remote_code: bool = False,
-    ):
-        self.quantize = quantize
-        self.process_group, rank, world_size = initialize_torch_distributed()
-        device = torch.device("hpu")
-        dtype = torch.bfloat16 if dtype is None else dtype
-        self.device, self.dtype = device, dtype
-
-        config = AutoConfig.from_pretrained(
-            model_id,
-            revision=revision,
-            trust_remote_code=trust_remote_code,
-        )
-        config.quantize = quantize
-        config.speculator = speculator
-        config.vision_config.quantize = quantize
-
-        tokenizer = AutoTokenizer.from_pretrained(
-            model_id,
-            revision=revision,
-            padding_side="left",
-            truncation_side="left",
-            trust_remote_code=trust_remote_code,
-        )
-        self.processor = AutoProcessor.from_pretrained(
-            model_id,
-            revision=revision,
-            padding_side="left",
-            truncation_side="left",
-            trust_remote_code=trust_remote_code,
-        )
-
-        weights_loader = get_loader(
-            quantize=quantize, model_id=model_id, revision=revision
-        )
-        torch.distributed.barrier(group=self.process_group)
-        filenames = weight_files(model_id, revision=revision, extension=".safetensors")
-        weights = Weights(
-            filenames,
-            device=device,
-            dtype=dtype,
-            process_group=self.process_group,
-            weights_loader=weights_loader,
-        )
-
-        model = IdeficsForVisionText2Text(config, weights)
-
-        self.config = config
-
-        torch.distributed.barrier(group=self.process_group)
-        super().__init__(
-            model_id=model_id,
-            model=model,
-            tokenizer=tokenizer,
-            requires_padding=True,
-            dtype=dtype,
-            device=device,
-            rank=rank,
-            world_size=world_size,
-        )
-
-    @property
-    def batch_type(self) -> Type[IdeficsCausalLMBatch]:
-        return IdeficsCausalLMBatch
-
-    def forward(
-        self,
-        input_ids,
-        attention_mask,
-        position_ids,
-        pixel_values,
-        image_hidden_states,
-        image_attention_mask,
-        past_key_values: Optional = None,
-    ) -> Tuple[torch.Tensor, List[Tuple[torch.Tensor, torch.Tensor]]]:
-        # Model Forward
-        kwargs = {
-            "input_ids": input_ids,
-            "attention_mask": attention_mask,
-            "pixel_values": pixel_values,
-            "image_hidden_states": image_hidden_states,
-            "image_attention_mask": image_attention_mask,
-            "past_key_values": past_key_values,
-            "use_cache": True,
-            "return_dict": True,
-        }
-        if self.has_position_ids:
-            kwargs["position_ids"] = position_ids
-
-        outputs, speculative_logits = self.model.forward(**kwargs)
-        return (
-            outputs.logits,
-            speculative_logits,
-            outputs.past_key_values,
-            outputs.image_hidden_states,
-        )
-
-    @tracer.start_as_current_span("generate_token")
-    def generate_token(
-        self, batch: IdeficsCausalLMBatch
-    ) -> Tuple[List[Generation], Optional[IdeficsCausalLMBatch], Tuple[int, int]]:
-        start = time.time_ns()
-        # slice the attention mask to the correct shape
-        attention_mask = batch.attention_mask[:, : -batch.padding_right_offset]
-        if batch.image_attention_mask is None:
-            image_attention_mask = None
-        else:
-            if batch.input_ids.size(1) == 1:
-                # THIS is a hack: when calling idefics.generate, the first time, we need the whole image_attention_mask (size bs x max_seq_len x max_num_images),
-                # but the subsequent times, we only need the last attention mask along the `max_seq_len` dimension
-                # this is due to the nature IDEFICS: it's an encoder decoder, and so when decoding, only the currently generated
-                # token need to attend to the encoder hidden states (i.e. the vision encoder)
-                # Also see seq2seq_lm.Seq2SeqLM.generate_token which has roughly the same logic
-                image_attention_mask = batch.image_attention_mask[
-                    :, -(batch.padding_right_offset + 1)
-                ].unsqueeze(1)
-            else:
-                image_attention_mask = batch.image_attention_mask[
-                    :, : -batch.padding_right_offset
-                ]
-
-        logits, speculative_logits, past, image_hidden_states = self.forward(
-            input_ids=batch.input_ids,
-            attention_mask=attention_mask,
-            position_ids=batch.position_ids,
-            pixel_values=batch.pixel_values,
-            image_hidden_states=batch.image_hidden_states,
-            image_attention_mask=image_attention_mask,
-            past_key_values=batch.past_key_values,
-        )
-        # Hardcoded remove image tokens
-        logits[:, 32000:32001] = torch.finfo(logits.dtype).min
-
-        start_decode = time.time_ns()
-
-        # Results
-        generations: List[Generation] = []
-        stopped = True
-
-        # Zipped iterator
-        iterator = zip(
-            batch.requests,
-            batch.input_lengths,
-            batch.prefix_offsets,
-            batch.read_offsets,
-            logits,
-            batch.next_token_choosers,
-            batch.stopping_criterias,
-            batch.all_input_ids,
-        )
-
-        # For each member of the batch
-        for i, (
-            request,
-            input_length,
-            prefix_offset,
-            read_offset,
-            logits,
-            next_token_chooser,
-            stopping_criteria,
-            all_input_ids,
-        ) in enumerate(iterator):
-            # Select next token
-            next_token_id, logprobs = next_token_chooser(
-                all_input_ids.view(1, -1), logits[-1:, :]
-            )
-
-            # Append next token to all tokens
-            all_input_ids = torch.cat([all_input_ids, next_token_id])
-            new_input_length = input_length + 1
-
-            # Generated token
-            next_token_logprob = logprobs[-1, next_token_id]
-            next_token_id_squeezed = next_token_id.squeeze()
-            next_token_text, prefix_offset, read_offset = self.decode_token(
-                all_input_ids[:, 0], prefix_offset, read_offset
-            )
-
-            # Evaluate stopping criteria
-            stop, reason = stopping_criteria(
-                next_token_id_squeezed,
-                next_token_text,
-            )
-
-            if not stop:
-                stopped = False
-
-            # Shard generations
-            # All generations will be appended in the rust sharded client
-            if i % self.world_size == self.rank:
-                if stop:
-                    # Decode generated tokens
-                    output_text, _, _ = self.decode_token(
-                        all_input_ids[:, 0],
-                        prefix_offset=len(all_input_ids)
-                        - stopping_criteria.current_tokens
-                        - 1,
-                        read_offset=len(all_input_ids)
-                        - stopping_criteria.current_tokens,
-                        skip_special_tokens=True,
-                    )
-                    # Get seed
-                    if isinstance(next_token_chooser.choice, Sampling):
-                        seed = next_token_chooser.choice.seed
-                    else:
-                        seed = None
-
-                    generated_text = GeneratedText(
-                        output_text, stopping_criteria.current_tokens, reason, seed
-                    )
-                else:
-                    generated_text = None
-
-                # Prefill
-                if stopping_criteria.current_tokens == 1 and request.prefill_logprobs:
-                    # Remove generated token to only have prefill and add nan for first prompt token
-                    prefill_logprobs = [float("nan")] + torch.log_softmax(
-                        logits, -1
-                    ).gather(1, all_input_ids[1:]).squeeze(1)[
-                        -new_input_length:-1
-                    ].tolist()
-                    prefill_token_ids = all_input_ids[-new_input_length:-1]
-                    prefill_texts = self.tokenizer.batch_decode(
-                        prefill_token_ids,
-                        clean_up_tokenization_spaces=False,
-                        skip_special_tokens=False,
-                    )
-                    prefill_tokens = Tokens(
-                        prefill_token_ids,
-                        prefill_logprobs,
-                        prefill_texts,
-                        is_special=[],
-                    )
-                else:
-                    prefill_tokens = None
-
-                top_tokens = None
-
-                generation = Generation(
-                    request.id,
-                    prefill_tokens,
-                    Tokens(
-                        [next_token_id_squeezed],
-                        [next_token_logprob],
-                        [next_token_text],
-                        [next_token_id_squeezed.item() in self.all_special_ids],
-                    ),
-                    generated_text,
-                    top_tokens,
-                )
-
-                generations.append(generation)
-
-            # Update values
-            batch.next_token_choosers[i] = batch.next_token_choosers[i].advance_grammar(
-                next_token_id_squeezed.item()
-            )
-            batch.input_ids[i, 0] = next_token_id
-            batch.all_input_ids[i] = all_input_ids
-            batch.input_lengths[i] = new_input_length
-            batch.prefix_offsets[i] = prefix_offset
-            batch.read_offsets[i] = read_offset
-            batch.max_input_length = max(batch.max_input_length, new_input_length)
-
-        # We finished all generations in the batch; there is no next batch
-        if stopped:
-            forward_ns = start_decode - start
-            decode_ns = time.time_ns() - start_decode
-            return generations, None, (forward_ns, decode_ns)
-
-        # Slice unused values from prefill
-        batch.input_ids = batch.input_ids[:, :1]
-
-        # Update attention_mask as we added a new token to input_ids
-        batch.attention_mask[:, -batch.padding_right_offset] = 1
-        batch.image_attention_mask[:, -batch.padding_right_offset, :] = (
-            batch.image_attention_mask[:, -(batch.padding_right_offset + 1), :]
-        )
-        # Decrease right offset
-        batch.padding_right_offset -= 1
-
-        # Update position_ids
-        batch.position_ids = batch.position_ids[:, -1:] + 1
-
-        # Update past key values
-        batch.past_key_values = past
-        batch.image_hidden_states = image_hidden_states
-
-        forward_ns = start_decode - start
-        decode_ns = time.time_ns() - start_decode
-        return generations, batch, (forward_ns, decode_ns)

backends/gaudi/server/text_generation_server/models/mamba.py

@@ -1,814 +0,0 @@
-import torch
-import torch.distributed
-from transformers import AutoTokenizer, PreTrainedTokenizerBase
-from typing import Optional
-from text_generation_server.models.custom_modeling.mamba_modeling import (
-    MambaConfig,
-)
-from loguru import logger
-from text_generation_server.pb import generate_pb2
-from text_generation_server.utils import (
-    initialize_torch_distributed,
-    weight_files,
-    Weights,
-)
-from text_generation_server.models.globals import CUDA_GRAPHS, MEM_POOL
-import time
-from text_generation_server.models.custom_modeling.mamba_modeling import (
-    MambaModel,
-    InferenceParams,
-)
-from text_generation_server.models import Model
-from typing import Any, List, Tuple, Type, Dict
-from text_generation_server.models.types import (
-    Batch,
-    Tokens,
-    Generation,
-    GeneratedText,
-)
-from text_generation_server.utils.chunks import concat_text_chunks
-from text_generation_server.utils.quantization import get_loader
-from text_generation_server.utils.tokens import batch_top_tokens, Sampling
-from dataclasses import dataclass
-from text_generation_server.utils import NextTokenChooser, StoppingCriteria
-
-
-def new_inference_params(
-    n_blocks: int,
-    batch_size: int,
-    d_inner: int,
-    d_conv: int,
-    d_state: int,
-    seqlen_offset: int,
-    dtype: torch.dtype,
-    device: torch.device,
-):
-    max_seqlen = 0
-    conv_states = torch.zeros(
-        (
-            n_blocks,
-            batch_size,
-            d_inner,
-            d_conv,
-        ),
-        device=device,
-        dtype=dtype,
-    )
-    ssm_states = torch.zeros(
-        (
-            n_blocks,
-            batch_size,
-            d_inner,
-            d_state,
-        ),
-        device=device,
-        dtype=dtype,
-    )
-    inference_params = InferenceParams(
-        max_seqlen=max_seqlen,
-        max_batch_size=batch_size,
-        seqlen_offset=seqlen_offset,
-        conv_states=conv_states,
-        ssm_states=ssm_states,
-    )
-    return inference_params
-
-
-@dataclass
-class MambaBatch(Batch):
-    batch_id: int
-    requests: List[generate_pb2.Request]
-    requests_idx_mapping: Dict[int, int]
-
-    # Decoder values
-    input_ids: torch.Tensor
-
-    # All tokens
-    all_input_ids: List[torch.Tensor]
-
-    # Lengths of all generations present in the batch
-    input_lengths: List[int]
-    prefix_offsets: List[int]
-    read_offsets: List[int]
-
-    # Generation helpers
-    next_token_choosers: List[NextTokenChooser]
-    stopping_criterias: List[StoppingCriteria]
-    top_n_tokens: List[int]
-    top_n_tokens_tensor: torch.Tensor
-
-    # Metadata used for padding
-    max_input_length: int
-    padding_right_offset: int
-
-    # Maximum number of tokens this batch will grow to
-    max_tokens: int
-
-    # Past metadata
-    keys_head_dim_last: bool = True
-
-    # Inference params
-    inference_params: Optional[Dict[str, Any]] = None
-
-    def to_pb(self) -> generate_pb2.CachedBatch:
-        return generate_pb2.CachedBatch(
-            id=self.batch_id,
-            request_ids=[r.id for r in self.requests],
-            size=len(self),
-            max_tokens=self.max_tokens,
-        )
-
-    @classmethod
-    def from_pb(
-        cls,
-        pb: generate_pb2.Batch,
-        tokenizer: PreTrainedTokenizerBase,
-        dtype: torch.dtype,
-        device: torch.device,
-    ) -> "MambaBatch":
-        inputs = []
-        next_token_choosers = []
-        stopping_criterias = []
-        top_n_tokens = []
-        prefix_offsets = []
-        read_offsets = []
-        requests_idx_mapping = {}
-
-        # Parse batch
-        max_truncation = 0
-        padding_right_offset = 0
-        max_decode_tokens = 0
-        for i, r in enumerate(pb.requests):
-            requests_idx_mapping[r.id] = i
-            inputs.append(concat_text_chunks(r.input_chunks.chunks))
-            next_token_choosers.append(
-                NextTokenChooser.from_pb(r.parameters, device, tokenizer)
-            )
-            stopping_criteria = StoppingCriteria.from_pb(
-                r.stopping_parameters, tokenizer
-            )
-            stopping_criterias.append(stopping_criteria)
-            top_n_tokens.append(r.top_n_tokens)
-            max_truncation = max(max_truncation, r.truncate)
-            max_decode_tokens += stopping_criteria.max_new_tokens
-            padding_right_offset = max(
-                padding_right_offset, stopping_criteria.max_new_tokens
-            )
-
-        tokenized_inputs = tokenizer(
-            inputs,
-            return_tensors="pt",
-            padding=True,
-            return_token_type_ids=False,
-            truncation=True,
-            max_length=max_truncation,
-        ).to(device)
-        for _ in pb.requests:
-            input_len = tokenized_inputs["input_ids"].shape[1]
-            prefix_offsets.append(input_len - 5)
-            read_offsets.append(input_len)
-
-        input_lengths = tokenized_inputs["attention_mask"].sum(1)
-        max_input_length = input_lengths.max()
-        input_ids = tokenized_inputs["input_ids"]
-        all_input_ids = tokenized_inputs["input_ids"].T.split(1, dim=1)
-        top_n_tokens_tensor = torch.tensor(
-            top_n_tokens, device=device, dtype=torch.int64
-        )
-        max_tokens = len(inputs) * (max_input_length + max_decode_tokens)
-        return cls(
-            batch_id=pb.id,
-            requests=pb.requests,
-            requests_idx_mapping=requests_idx_mapping,
-            input_ids=input_ids,
-            # past_input_ids=None,
-            all_input_ids=list(all_input_ids),
-            input_lengths=input_lengths.tolist(),
-            prefix_offsets=prefix_offsets,
-            read_offsets=read_offsets,
-            next_token_choosers=next_token_choosers,
-            stopping_criterias=stopping_criterias,
-            top_n_tokens=top_n_tokens,
-            top_n_tokens_tensor=top_n_tokens_tensor,
-            max_input_length=max_input_length.item(),
-            padding_right_offset=padding_right_offset,
-            max_tokens=max_tokens,
-        )
-
-    def filter(self, request_ids: List[int]) -> Optional["MambaBatch"]:
-        if len(request_ids) == 0:
-            raise ValueError("Batch must have at least one request")
-        if len(request_ids) == len(self):
-            return self
-
-        keep_indices = []
-
-        # New values after filtering
-        requests_idx_mapping = {}
-        requests = []
-        input_lengths = []
-        prefix_offsets = []
-        read_offsets = []
-        all_input_ids = []
-        max_input_length = 0
-
-        next_token_choosers = []
-        stopping_criterias = []
-        top_n_tokens = []
-
-        total_remaining_decode_tokens = 0
-        new_padding_right_offset = 0
-
-        indices = []
-        for i, request_id in enumerate(request_ids):
-            idx = self.requests_idx_mapping[request_id]
-            requests_idx_mapping[request_id] = i
-            keep_indices.append(idx)
-
-            requests.append(self.requests[idx])
-            prefix_offsets.append(self.prefix_offsets[idx])
-            read_offsets.append(self.read_offsets[idx])
-            all_input_ids.append(self.all_input_ids[idx])
-
-            request_input_length = self.input_lengths[idx]
-            input_lengths.append(request_input_length)
-            max_input_length = max(max_input_length, request_input_length)
-            indices.append(idx)
-
-            next_token_choosers.append(self.next_token_choosers[idx])
-            stopping_criteria = self.stopping_criterias[idx]
-            stopping_criterias.append(stopping_criteria)
-            top_n_tokens.append(self.top_n_tokens[idx])
-            remaining_decode_tokens = (
-                stopping_criteria.max_new_tokens - stopping_criteria.current_tokens
-            )
-            total_remaining_decode_tokens += remaining_decode_tokens
-            new_padding_right_offset = max(
-                new_padding_right_offset, remaining_decode_tokens
-            )
-
-        # Apply indices to input_ids, attention mask, past key values and other items that need to be cached
-        input_ids = self.input_ids[keep_indices]
-
-        top_n_tokens_tensor = self.top_n_tokens_tensor[keep_indices]
-        max_tokens = len(request_ids) * max_input_length + total_remaining_decode_tokens
-
-        self.requests = requests
-        self.requests_idx_mapping = requests_idx_mapping
-        self.input_ids = input_ids
-        self.all_input_ids = all_input_ids
-        self.input_lengths = input_lengths
-        self.prefix_offsets = prefix_offsets
-        self.read_offsets = read_offsets
-        self.next_token_choosers = next_token_choosers
-        self.stopping_criterias = stopping_criterias
-        self.top_n_tokens = top_n_tokens
-        self.top_n_tokens_tensor = top_n_tokens_tensor
-        self.max_input_length = max_input_length
-        self.padding_right_offset = new_padding_right_offset
-        self.max_tokens = max_tokens
-
-        # TODO
-        # Kept it simple by just updating the state, maybe updating the other CPU values is necessary.
-        self.inference_params.conv_states = self.inference_params.conv_states[
-            :, indices
-        ]
-        self.inference_params.ssm_states = self.inference_params.ssm_states[:, indices]
-        return self
-
-    @classmethod
-    def concatenate(cls, batches: List["MambaBatch"]) -> "MambaBatch":
-        # Used for padding
-        total_batch_size = 0
-        max_input_length = 0
-        padding_right_offset = 0
-        for batch in batches:
-            total_batch_size += len(batch)
-            max_input_length = max(max_input_length, batch.max_input_length)
-            padding_right_offset = max(padding_right_offset, batch.padding_right_offset)
-
-        # Batch attributes
-        requests = []
-        requests_idx_mapping = {}
-        input_lengths = []
-        prefix_offsets = []
-        read_offsets = []
-        all_input_ids = []
-        next_token_choosers = []
-        stopping_criterias = []
-        top_n_tokens = []
-        max_tokens = 0
-        seqlen_offset = 0
-
-        (n_blocks, _, d_inner, d_conv) = batches[0].inference_params.conv_states.shape
-        (_, _, _, d_state) = batches[0].inference_params.ssm_states.shape
-        dtype = batches[0].inference_params.conv_states.dtype
-        device = batches[0].inference_params.conv_states.device
-        inference_params = new_inference_params(
-            n_blocks=n_blocks,
-            batch_size=total_batch_size,
-            d_state=d_state,
-            d_conv=d_conv,
-            d_inner=d_inner,
-            seqlen_offset=seqlen_offset,
-            device=device,
-            dtype=dtype,
-        )
-
-        # Batch tensors
-        input_ids = None
-        top_n_tokens_tensor = None
-
-        # Used for slicing correctly inside the tensors
-        # Equivalent to a cumsum on batch sizes
-        start_index = 0
-        for i, batch in enumerate(batches):
-            requests.extend(batch.requests)
-            input_lengths.extend(batch.input_lengths)
-            prefix_offsets.extend(batch.prefix_offsets)
-            read_offsets.extend(batch.read_offsets)
-            all_input_ids.extend(batch.all_input_ids)
-            next_token_choosers.extend(batch.next_token_choosers)
-            stopping_criterias.extend(batch.stopping_criterias)
-            top_n_tokens.extend(batch.top_n_tokens)
-
-            if i == 0:
-                requests_idx_mapping = batch.requests_idx_mapping
-            else:
-                # We need to offset the mapping for each batch by the cumulative batch size
-                for k, v in batch.requests_idx_mapping.items():
-                    requests_idx_mapping[k] = v + start_index
-
-            # Slicing end index for this batch
-            end_index = start_index + len(batch)
-
-            # Create empty tensor
-            # input_ids is always of shape [batch_size, 1]
-            # We do not need to pad it
-            if input_ids is None:
-                input_ids = batch.input_ids.new_empty((total_batch_size, 1))
-            # Copy to correct indices
-            input_ids[start_index:end_index] = batch.input_ids
-
-            if top_n_tokens_tensor is None:
-                top_n_tokens_tensor = batches[0].top_n_tokens_tensor.new_zeros(
-                    total_batch_size,
-                )
-            top_n_tokens_tensor[start_index:end_index] = batch.top_n_tokens_tensor
-
-            # Add eventual padding tokens that were added while concatenating
-            max_tokens += batch.max_tokens + (
-                max_input_length - batch.max_input_length
-            ) * len(batch)
-
-            inference_params.max_seqlen = max(
-                inference_params.max_seqlen, batch.inference_params.max_seqlen
-            )
-            assert batch.inference_params.seqlen_offset != 0, "Invalid seqlen offset"
-            inference_params.seqlen_offset = max(
-                inference_params.seqlen_offset, batch.inference_params.seqlen_offset
-            )
-
-            inference_params.conv_states[:, start_index:end_index] = (
-                batch.inference_params.conv_states
-            )
-            inference_params.ssm_states[:, start_index:end_index] = (
-                batch.inference_params.ssm_states
-            )
-
-            start_index = end_index
-
-        return cls(
-            batch_id=batches[0].batch_id,
-            requests=requests,
-            requests_idx_mapping=requests_idx_mapping,
-            input_ids=input_ids,
-            all_input_ids=all_input_ids,
-            input_lengths=input_lengths,
-            prefix_offsets=prefix_offsets,
-            read_offsets=read_offsets,
-            next_token_choosers=next_token_choosers,
-            stopping_criterias=stopping_criterias,
-            top_n_tokens=top_n_tokens,
-            top_n_tokens_tensor=top_n_tokens_tensor,
-            max_input_length=max_input_length,
-            padding_right_offset=padding_right_offset,
-            keys_head_dim_last=batches[0].keys_head_dim_last,
-            max_tokens=max_tokens,
-            inference_params=inference_params,
-        )
-
-    def __len__(self):
-        return len(self.requests)
-
-
-class Mamba(Model):
-    def __init__(
-        self,
-        model_id: str,
-        revision: Optional[str] = None,
-        quantize: Optional[str] = None,
-        speculator: Optional[str] = None,
-        dtype: Optional[torch.dtype] = None,
-        trust_remote_code: bool = False,
-    ):
-        self.quantize = quantize
-        self.process_group, _rank, world_size = initialize_torch_distributed()
-        if world_size > 1:
-            raise RuntimeError("Mamba does not support Tensor Parallelism (TP)")
-        self.cuda_graphs = {}
-        if torch.cuda.is_available():
-            device = torch.device("cuda")
-            # Bf16 is important. In f16 accumulations in the matmul are causing
-            # differences while the server is under load.
-            # This is detectable by the integration load test
-            dtype = torch.bfloat16 if dtype is None else dtype
-        else:
-            if quantize:
-                raise ValueError("quantization is not available on CPU")
-
-            device = torch.device("cpu")
-            dtype = torch.float32 if dtype is None else dtype
-
-        tokenizer = AutoTokenizer.from_pretrained(
-            "EleutherAI/gpt-neox-20b",
-            revision=revision,
-            padding_side="left",
-            truncation_side="left",
-            trust_remote_code=trust_remote_code,
-        )
-        config = MambaConfig.from_pretrained(
-            model_id, revision=revision, trust_remote_code=trust_remote_code
-        )
-
-        tokenizer.bos_token_id = config.bos_token_id
-        tokenizer.eos_token_id = config.eos_token_id
-        tokenizer.pad_token = tokenizer.eos_token
-
-        config.quantize = quantize
-        config.speculator = speculator
-        torch.distributed.barrier(group=self.process_group)
-        weights_loader = get_loader(
-            quantize=quantize, model_id=model_id, revision=revision
-        )
-        filenames = weight_files(model_id, revision=revision, extension=".safetensors")
-        weights = Weights(
-            filenames,
-            device,
-            dtype,
-            process_group=self.process_group,
-            weights_loader=weights_loader,
-        )
-        model = MambaModel(config, weights)
-        torch.distributed.barrier(group=self.process_group)
-        super(Mamba, self).__init__(
-            model_id=model_id,
-            model=model,
-            tokenizer=tokenizer,
-            requires_padding=True,
-            dtype=dtype,
-            device=device,
-        )
-
-    @property
-    def batch_type(self) -> Type[MambaBatch]:
-        return MambaBatch
-
-    def warmup(self, batch) -> Optional[int]:
-        # TODO: implement warmup for Mamba if needed
-        if CUDA_GRAPHS:
-            if self.speculate is None or self.speculate == 0:
-                try:
-                    logger.info(f"Cuda Graphs are enabled for sizes {CUDA_GRAPHS}")
-                    # Warmup cuda graphs
-                    for bs in CUDA_GRAPHS:
-                        self.cuda_graph_warmup(bs)
-                except Exception:
-                    logger.exception("Decode cuda graph warmup failed")
-        else:
-            logger.info(f"Cuda Graphs are disabled (CUDA_GRAPHS={CUDA_GRAPHS}).")
-
-        return None
-
-    def cuda_graph_warmup(self, batch_size: int):
-        input_ids = torch.zeros((batch_size, 1), dtype=torch.int64, device=self.device)
-        n_blocks = len(self.model.blocks)
-
-        d_state = self.model.config.d_state
-        d_conv = self.model.config.d_conv
-        # Inner takes the expand multiplication
-        d_inner = self.model.config.d_inner
-
-        # Important seqlen_offset to go through the update mecanism with the state
-        seqlen_offset = 1
-        inference_params = new_inference_params(
-            n_blocks=n_blocks,
-            batch_size=batch_size,
-            d_state=d_state,
-            d_conv=d_conv,
-            d_inner=d_inner,
-            seqlen_offset=seqlen_offset,
-            device=self.device,
-            dtype=self.dtype,
-        )
-
-        graph = torch.cuda.CUDAGraph()
-
-        torch.cuda.synchronize()
-        # Run once outside to warmup
-        self.model.forward(input_ids=input_ids, inference_params=inference_params)
-        torch.cuda.synchronize()
-
-        with torch.cuda.graph(graph, pool=MEM_POOL):
-            logits, speculative_logits = self.model.forward(
-                input_ids=input_ids, inference_params=inference_params
-            )
-        torch.cuda.synchronize()
-        graph_dict = {
-            "input_ids": input_ids,
-            "inference_params": inference_params,
-            "graph": graph,
-            "logits": logits,
-            "speculative_logits": speculative_logits,
-        }
-        self.cuda_graphs[batch_size] = graph_dict
-
-    def tunableop_warmup(self, batch_size: int, seqlen: int):
-        input_ids = torch.zeros((batch_size, 1), dtype=torch.int64, device=self.device)
-        n_blocks = len(self.model.blocks)
-
-        d_state = self.model.config.d_state
-        d_conv = self.model.config.d_conv
-        # Inner takes the expand multiplication
-        d_inner = self.model.config.d_inner
-
-        # Important seqlen_offset to go through the update mecanism with the state
-        seqlen_offset = 1
-        inference_params = new_inference_params(
-            n_blocks=n_blocks,
-            batch_size=seqlen,
-            d_state=d_state,
-            d_conv=d_conv,
-            d_inner=d_inner,
-            seqlen_offset=seqlen_offset,
-            device=self.device,
-            dtype=self.dtype,
-        )
-
-        self.model.forward(input_ids=input_ids, inference_params=inference_params)
-
-    def forward(
-        self, input_ids: torch.Tensor, inference_params: Any
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        bs = input_ids.shape[0]
-        padded_bs = bs
-        if bs == 3:
-            padded_bs = 4
-        elif 3 < bs <= 8:
-            padded_bs = 8
-        elif bs > 8:
-            padded_bs = (bs + 7) // 8 * 8
-
-        # Try to find an associated cuda graph
-        cuda_graph = self.cuda_graphs.get(padded_bs, None)
-        is_prefill = inference_params is None or inference_params.seqlen_offset == 0
-
-        if is_prefill or cuda_graph is None:
-            return self.model(
-                input_ids,
-                inference_params=inference_params,
-            )
-
-        # Copy inputs to the static inputs of the cuda graph
-        # Static inputs are potentially padded
-        cuda_graph["input_ids"][:bs] = input_ids
-        cuda_graph["inference_params"].conv_states[
-            :, :bs
-        ] = inference_params.conv_states
-        cuda_graph["inference_params"].ssm_states[:, :bs] = inference_params.ssm_states
-
-        # Replay the graph
-        cuda_graph["graph"].replay()
-
-        inference_params.conv_states.copy_(
-            cuda_graph["inference_params"].conv_states[:, :bs]
-        )
-        inference_params.ssm_states.copy_(
-            cuda_graph["inference_params"].ssm_states[:, :bs]
-        )
-        # Slice output to the correct shape
-        speculative_logits = (
-            cuda_graph["speculative_logits"][:bs]
-            if cuda_graph["speculative_logits"] is not None
-            else None
-        )
-        logits = cuda_graph["logits"][:bs]
-        return logits, speculative_logits
-
-    def generate_token(self, batch) -> Tuple[List[Any], Optional[Any], Tuple[int, int]]:
-        start = time.time_ns()
-        input_ids = (
-            batch.input_ids
-        )  # batch.past_input_ids if batch.past_input_ids is not None else batch.input_ids
-
-        batch_size, max_seqlen = input_ids.shape
-        # Inference params
-
-        if batch.inference_params is None:
-            # 0 is important here
-            seqlen_offset = 0
-            n_blocks = len(self.model.blocks)
-            d_state = self.model.config.d_state
-            d_conv = self.model.config.d_conv
-            d_inner = self.model.config.d_inner
-            inference_params = new_inference_params(
-                n_blocks=n_blocks,
-                batch_size=batch_size,
-                d_state=d_state,
-                d_conv=d_conv,
-                d_inner=d_inner,
-                seqlen_offset=seqlen_offset,
-                device=self.device,
-                dtype=self.dtype,
-            )
-            batch.inference_params = inference_params
-
-        # Forward pass
-        logits, speculative_logits = self.forward(
-            input_ids, inference_params=batch.inference_params
-        )
-
-        # batch.inference_params = new_inference_params
-        # Results
-        generations: List[Generation] = []
-        stopped = True
-
-        # Speculation is not active for causal
-        accepted_ids = torch.ones_like(batch.input_ids)[:, 0]
-        batch_top_token_ids, batch_top_token_logprobs = batch_top_tokens(
-            batch.top_n_tokens,
-            batch.top_n_tokens_tensor,
-            torch.log_softmax(logits[:, -1], -1),
-            accepted_ids,
-        )
-
-        start_decode = time.time_ns()
-
-        # Zipped iterator
-        iterator = zip(
-            batch.requests,
-            batch.input_lengths,
-            batch.prefix_offsets,
-            batch.read_offsets,
-            logits,
-            batch.next_token_choosers,
-            batch.stopping_criterias,
-            batch.all_input_ids,
-            batch.top_n_tokens,
-            batch_top_token_ids,
-            batch_top_token_logprobs,
-        )
-
-        # For each member of the batch
-        for i, (
-            request,
-            input_length,
-            prefix_offset,
-            read_offset,
-            logits,
-            next_token_chooser,
-            stopping_criteria,
-            all_input_ids,
-            top_n_tokens,
-            top_token_ids,
-            top_token_logprobs,
-        ) in enumerate(iterator):
-            # Select next token
-            next_token_id, logprobs = next_token_chooser(
-                all_input_ids.view(1, -1), logits[-1:, :]
-            )
-
-            # Append next token to all tokens
-            all_input_ids = torch.cat([all_input_ids, next_token_id])
-            new_input_length = input_length + 1
-
-            # Generated token
-            next_token_logprob = logprobs[-1, next_token_id]
-            next_token_id_squeezed = next_token_id.squeeze()
-            next_token_text, prefix_offset, read_offset = self.decode_token(
-                all_input_ids[:, 0], prefix_offset, read_offset
-            )
-
-            # Evaluate stopping criteria
-            stop, reason = stopping_criteria(
-                next_token_id_squeezed,
-                next_token_text,
-            )
-
-            if not stop:
-                stopped = False
-
-            # Shard generations
-            # All generations will be appended in the rust sharded client
-            if i % self.world_size == self.rank:
-                if stop:
-                    # Decode generated tokens
-                    output_text, _, _ = self.decode_token(
-                        all_input_ids[:, 0],
-                        prefix_offset=len(all_input_ids)
-                        - stopping_criteria.current_tokens
-                        - 1,
-                        read_offset=len(all_input_ids)
-                        - stopping_criteria.current_tokens,
-                        skip_special_tokens=True,
-                    )
-                    # Get seed
-                    if isinstance(next_token_chooser.choice, Sampling):
-                        seed = next_token_chooser.choice.seed
-                    else:
-                        seed = None
-
-                    generated_text = GeneratedText(
-                        output_text, stopping_criteria.current_tokens, reason, seed
-                    )
-                else:
-                    generated_text = None
-
-                if stopping_criteria.current_tokens == 1 and request.prefill_logprobs:
-                    # Remove generated token to only have prefill and add nan for first prompt token
-                    prefill_logprobs = [float("nan")] + torch.log_softmax(
-                        logits, -1
-                    ).gather(1, all_input_ids[1:]).squeeze(1)[
-                        -new_input_length:-1
-                    ].tolist()
-                    prefill_token_ids = all_input_ids[-new_input_length:-1]
-                    prefill_texts = self.tokenizer.batch_decode(
-                        prefill_token_ids,
-                        clean_up_tokenization_spaces=False,
-                        skip_special_tokens=False,
-                    )
-                    prefill_tokens = Tokens(
-                        prefill_token_ids,
-                        prefill_logprobs,
-                        prefill_texts,
-                        is_special=[],
-                    )
-                else:
-                    prefill_tokens = None
-
-                if top_n_tokens > 0:
-                    toptoken_texts = self.tokenizer.batch_decode(
-                        top_token_ids,
-                        clean_up_tokenization_spaces=False,
-                        skip_special_tokens=False,
-                    )
-                    special_toptokens = [
-                        token_id in self.all_special_ids for token_id in top_token_ids
-                    ]
-                    top_tokens = Tokens(
-                        top_token_ids,
-                        top_token_logprobs,
-                        toptoken_texts,
-                        special_toptokens,
-                    )
-                else:
-                    top_tokens = None
-
-                generation = Generation(
-                    request.id,
-                    prefill_tokens,
-                    Tokens(
-                        [next_token_id_squeezed],
-                        [next_token_logprob],
-                        [next_token_text],
-                        [next_token_id_squeezed.item() in self.all_special_ids],
-                    ),
-                    generated_text,
-                    top_tokens,
-                )
-
-                generations.append(generation)
-
-                # Update values
-                batch.next_token_choosers[i] = batch.next_token_choosers[
-                    i
-                ].advance_grammar(next_token_id_squeezed.item())
-                batch.input_ids[i, 0] = next_token_id
-                batch.all_input_ids[i] = all_input_ids
-                batch.input_lengths[i] = new_input_length
-                batch.prefix_offsets[i] = prefix_offset
-                batch.read_offsets[i] = read_offset
-                batch.max_input_length = max(batch.max_input_length, new_input_length)
-
-        # We finished all generations in the batch; there is no next batch
-        if stopped:
-            forward_ns = start_decode - start
-            decode_ns = time.time_ns() - start_decode
-            return generations, None, (forward_ns, decode_ns)
-
-        # Slice unused values from prefill
-        batch.input_ids = batch.input_ids[:, :1]
-
-        forward_ns = start_decode - start
-        decode_ns = time.time_ns() - start_decode
-        return generations, batch, (forward_ns, decode_ns)

backends/gaudi/server/text_generation_server/models/starcoder.py

@@ -1,47 +0,0 @@
-import torch
-from dataclasses import dataclass
-from typing import List, Optional, Type
-
-from text_generation_server.models import CausalLM
-from text_generation_server.models.causal_lm import CausalLMBatch
-
-
-@dataclass
-class StarCoderCausalLMBatch(CausalLMBatch):
-    past_key_values: Optional[List[torch.Tensor]]
-
-    def detach_kv_cache(self):
-        past_keys = []
-        past_values = []
-        last_dim = int(self.past_key_values[0].size(dim=-1) / 2)
-        for key_value in self.past_key_values:
-            past_keys.append(key_value.split((last_dim, last_dim), dim=-1)[0])
-            past_values.append(key_value.split((last_dim, last_dim), dim=-1)[1])
-        del self.past_key_values
-
-        return past_keys, past_values
-
-    def attach_kv_cache(self, past_keys, past_values):
-        self.past_key_values = [
-            torch.cat((key, value), dim=-1)
-            for key, value in zip(past_keys, past_values)
-        ]
-
-
-class StarCoder(CausalLM):
-    def __init__(
-        self,
-        model_id: str,
-        revision: Optional[str] = None,
-        dtype: Optional[torch.dtype] = None,
-    ):
-
-        super(StarCoder, self).__init__(
-            model_id=model_id,
-            revision=revision,
-            dtype=dtype,
-        )
-
-    @property
-    def batch_type(self) -> Type[CausalLMBatch]:
-        return StarCoderCausalLMBatch

backends/gaudi/tgi-entrypoint.sh

@@ -7,13 +7,5 @@ if [[ "$*" == *"--sharded true"* ]]; then
   echo 'setting PT_HPU_ENABLE_LAZY_COLLECTIVES=1 for sharding'
   export PT_HPU_ENABLE_LAZY_COLLECTIVES=1
 fi
-# Check if ATTENTION environment variable is set to paged
-if [[ "$ATTENTION" == "paged" ]]; then
-  # Check if Llama-4 is in the command line arguments
-  if [[ "$*" == *"Llama-4"* || "$*" == *"Qwen3"* ]]; then
-    echo 'ATTENTION=paged and Llama-4 or Qwen3 detected'
-    pip install git+https://github.com/huggingface/transformers.git@29338949
-  fi
-fi
 
 text-generation-launcher $@