perfect the adaptation of v3.0.0

server/exllamav2_kernels/out/special_tokens_map.json

+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

server/exllamav2_kernels/out/tokenizer_config.json

+{
+  "add_bos_token": true,
+  "add_eos_token": false,
+  "add_prefix_space": null,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "legacy": true,
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": null,
+  "sp_model_kwargs": {},
+  "tokenizer_class": "LlamaTokenizer",
+  "unk_token": "<unk>",
+  "use_default_system_prompt": false
+}

server/out/special_tokens_map.json

+{
+  "bos_token": {
+    "content": "<s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "</s>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<unk>",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  }
+}

server/out/tokenization_baichuan.py

+# Copyright 2023 Baichuan Inc. All Rights Reserved.
+
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "tokenizer.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {},
+    "tokenizer_file": {},
+}
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {}
+
+
+class BaichuanTokenizer(PreTrainedTokenizer):
+    """
+    Construct a Baichuan tokenizer. Based on byte-level Byte-Pair-Encoding.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        unk_token="<unk>",
+        bos_token="<s>",
+        eos_token="</s>",
+        pad_token=None,
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        add_bos_token=True,
+        add_eos_token=False,
+        clean_up_tokenization_spaces=False,
+        **kwargs,
+    ):
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+        bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
+        eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
+        unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
+        pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
+        self.vocab_file = vocab_file # shantf 12-11号改动
+        self.add_bos_token = add_bos_token
+        self.add_eos_token = add_eos_token
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(vocab_file)
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            add_bos_token=add_bos_token,
+            add_eos_token=add_eos_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(self.vocab_file)
+
+    @property
+    def vocab_size(self):
+        """Returns vocab size"""
+        return self.sp_model.get_piece_size()
+
+    def get_vocab(self):
+        """Returns vocab as a dict"""
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        """Returns a tokenized string."""
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        token = self.sp_model.IdToPiece(index)
+        return token
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (string) in a single string."""
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for i, token in enumerate(tokens):
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special and i != 0:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (`str`):
+                The directory in which to save the vocabulary.
+
+        Returns:
+            `Tuple(str)`: Paths to the files saved.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = bos_token_id + token_ids_0 + eos_token_id
+
+        if token_ids_1 is not None:
+            output = output + bos_token_id + token_ids_1 + eos_token_id
+
+        return output
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        bos_token_id = [1] if self.add_bos_token else []
+        eos_token_id = [1] if self.add_eos_token else []
+
+        if token_ids_1 is None:
+            return bos_token_id + ([0] * len(token_ids_0)) + eos_token_id
+        return (
+            bos_token_id
+            + ([0] * len(token_ids_0))
+            + eos_token_id
+            + bos_token_id
+            + ([0] * len(token_ids_1))
+            + eos_token_id
+        )
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Creates a mask from the two sequences passed to be used in a sequence-pair classification task. An ALBERT
+        sequence pair mask has the following format:
+
+        ```
+        0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+        | first sequence    | second sequence |
+        ```
+
+        if token_ids_1 is None, only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of ids.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [token type IDs](../glossary#token-type-ids) according to the given sequence(s).
+        """
+        bos_token_id = [self.bos_token_id] if self.add_bos_token else []
+        eos_token_id = [self.eos_token_id] if self.add_eos_token else []
+
+        output = [0] * len(bos_token_id + token_ids_0 + eos_token_id)
+
+        if token_ids_1 is not None:
+            output += [1] * len(bos_token_id + token_ids_1 + eos_token_id)
+
+        return output

server/out/tokenizer_config.json

+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<unk>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "</s>",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "auto_map": {
+    "AutoTokenizer": [
+      "tokenization_baichuan.BaichuanTokenizer",
+      null
+    ]
+  },
+  "bos_token": "<s>",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "</s>",
+  "model_max_length": 4096,
+  "pad_token": "<unk>",
+  "sp_model_kwargs": {},
+  "tokenizer_class": "BaichuanTokenizer",
+  "unk_token": "<unk>",
+  "use_fast": false
+}

server/text_generation_server/layers/attention/rocm.py

@@ -19,7 +19,8 @@ ENGINE = "triton" if use_triton else "ck"
 use_rocm_custom_paged_attn = os.getenv("ROCM_USE_CUSTOM_PAGED_ATTN", "1") != "0"
 try:
     if use_rocm_custom_paged_attn:
-        from vllm._custom_C import paged_attention_custom
+        #from vllm._custom_C import paged_attention_custom
+        from vllm import _custom_ops
 except ImportError as e:
     log_master(
         logger.info,
@@ -79,6 +80,7 @@ def paged_attention(
         _PARTITION_SIZE = _PARTITION_SIZE_V1V2
     else:
         _PARTITION_SIZE = _PARTITION_SIZE_CUSTOM
+    _PARTITION_SIZE = 512
     
     max_num_partitions = (max_s + _PARTITION_SIZE - 1) // _PARTITION_SIZE
     input_lengths = seqlen.input_lengths + seqlen.cache_lengths
@@ -103,7 +105,8 @@ def paged_attention(
             query,
             kv_cache.key,
             kv_cache.value,
-            kv_head_mapping,
+            #kv_head_mapping,
+            kv_head_mapping.shape[0],
             softmax_scale,
             block_tables,
             input_lengths,
@@ -128,7 +131,7 @@ def paged_attention(
         )
         max_logits = torch.empty_like(exp_sums)
 
-        if not use_custom:
+        #if not use_custom:
         ops.paged_attention_v2(
             out,
             exp_sums,
@@ -137,7 +140,8 @@ def paged_attention(
             query,
             kv_cache.key,
             kv_cache.value,
-                kv_head_mapping,
+            #kv_head_mapping,
+            kv_head_mapping.shape[0],
             softmax_scale,
             block_tables,
             input_lengths,
@@ -147,24 +151,24 @@ def paged_attention(
             "auto",
             1.0,
         )
-        else:
-            paged_attention_custom(
-                out,
-                exp_sums,
-                max_logits,
-                tmp_output,
-                query,
-                kv_cache.key,
-                kv_cache.value,
-                num_kv_heads,
-                softmax_scale,
-                block_tables,
-                input_lengths,
-                block_size,
-                max_s,
-                None,
-                "auto",
-            )
+        # else:
+        #     paged_attention_custom(
+        #         out,
+        #         exp_sums,
+        #         max_logits,
+        #         tmp_output,
+        #         query,
+        #         kv_cache.key,
+        #         kv_cache.value,
+        #         num_kv_heads,
+        #         softmax_scale,
+        #         block_tables,
+        #         input_lengths,
+        #         block_size,
+        #         max_s,
+        #         None,
+        #         "auto",
+        #     )
 
     return out
 

server/text_generation_server/layers/gptq/exllamav2.py

@@ -12,10 +12,11 @@ from text_generation_server.layers.gptq import GPTQWeight
 from text_generation_server.utils.log import log_master
 
 try:
-    from exllamav2.ext import exllamav2_ext
+    # from exllamav2.ext import exllamav2_ext
 
-    make_q_matrix = exllamav2_ext.make_q_matrix
-    gemm_half_q_half = exllamav2_ext.gemm_half_q_half
+    # make_q_matrix = exllamav2_ext.make_q_matrix
+    # gemm_half_q_half = exllamav2_ext.gemm_half_q_half
+    from exllamav2_kernels import make_q_matrix, gemm_half_q_half
 except ImportError:
     log_master(logger.warning, "exllamav2_kernels not installed.")
     raise
@@ -122,9 +123,9 @@ def ext_make_q_matrix(
                 w.qzeros,
                 w.scales,
                 w.g_idx.cpu(),
-                none_tensor,  # bias
+                #none_tensor,  # bias
                 temp_dq,
-                max_dq_rows,
+                #max_dq_rows,
             )
         # GPTQ without g_idx
         else:
@@ -139,9 +140,9 @@ def ext_make_q_matrix(
                 w.qzeros,
                 w.scales,
                 none_tensor,  # g_idx
-                none_tensor,  # bias
+                #none_tensor,  # bias
                 temp_dq,
-                max_dq_rows,
+                #max_dq_rows,
             )
     else:
         RuntimeError("Cannot create handle")

server/text_generation_server/layers/layernorm.py

@@ -172,7 +172,7 @@ class FastRMSNorm(nn.Module):
             residual = hidden_states
 
             out = torch.empty_like(hidden_states)
-            ops.rms_norm(
+            _custom_ops.rms_norm(
                 out,
                 hidden_states,
                 self.weight.data,

server/text_generation_server/layers/linear.py

@@ -8,7 +8,7 @@ import os
 #         "true",
 #         "1",
 #     )
-
+ROCM_USE_SKINNY_GEMM=False
 #     if ROCM_USE_SKINNY_GEMM:
 #         try:
 #             from vllm import _custom_C

server/text_generation_server/layers/rotary.py

@@ -7,7 +7,7 @@ from text_generation_server.utils.import_utils import SYSTEM
 if SYSTEM == "cuda":
     import rotary_emb
 elif SYSTEM == "rocm":
-    from vllm._C import ops
+    import vllm._custom_ops as ops
 elif SYSTEM == "ipex":
     import intel_extension_for_pytorch as ipex
 
@@ -67,7 +67,8 @@ class PositionRotaryEmbedding(nn.Module):
             head_size = query.shape[-1]
 
             # Inplace operation, updating query and key.
-            ops.rotary_embedding(query, key, head_size, cos, sin, True)
+            #ops.rotary_embedding(query, key, head_size, cos, sin, True)
+            torch.ops._C.rotary_embedding_tgi(query, key, head_size, cos, sin, True)
         elif SYSTEM == "ipex":
             ipex.llm.functional.rotary_embedding(
                 query, key, sin, cos, query.size(-1), True

server/text_generation_server/models/custom_modeling/flash_cohere_modeling.py

@@ -75,7 +75,7 @@ class CohereRotary(PositionRotaryEmbedding):
 
             rotary_emb.apply_rotary(k1, k2, cos, sin, k1, k2, False)
         elif SYSTEM == "rocm":
-            from vllm._C import ops
+            import vllm._custom_ops as ops
 
             # NOTE: On RoCm systems, we use a ROPE implementatation adapted from VLLM which launches a single kernel for both query/key, contrary to flash-attn implementation used on NVIDIA systems.
             # Compiling flash-attn rotary on RoCm, it appears hipcc is unable to unroll loops, resulting in an even slower inference compared to eager: https://github.com/pytorch/pytorch/issues/113773
@@ -83,7 +83,7 @@ class CohereRotary(PositionRotaryEmbedding):
             head_size = query.shape[-1]
 
             # Inplace operation, updating query and key.
-            ops.rotary_embedding(query, key, head_size, cos, sin, False)
+            torch.ops._C.rotary_embedding_tgi(query, key, head_size, cos, sin, True)
         elif SYSTEM == "ipex":
             import intel_extension_for_pytorch as ipex
 

server/text_generation_server/models/custom_modeling/flash_deepseek_v2_modeling.py

@@ -43,7 +43,7 @@ from text_generation_server.utils.weights import Weights
 
 if SYSTEM == "rocm":
     try:
-        from vllm import _custom_C
+        from vllm import _custom_ops
     except Exception as e:
         raise ImportError(f"Could not load `vllm._custom_C`. Full error: {e}")
 
@@ -395,22 +395,22 @@ class DeepseekV2MLP(nn.Module):
         self.quantize = config.quantize
 
     def forward(self, hidden_states: torch.Tensor, reduce: bool = True):
-        if (
-            SYSTEM == "rocm"
-            and self.hidden_act == "silu"
-            and hidden_states.dtype == torch.float16
-            and hidden_states.shape[0] == 1
-            and not self.quantize
-        ):
-            out = torch.empty(
-                hidden_states.shape[0],
-                self.intermediate_size,
-                dtype=hidden_states.dtype,
-                device="cuda",
-            )
-            _custom_C.LLMM_Silu(self.gate_up_proj.linear.weight, hidden_states, out, 8)
-            return self.down_proj(out, reduce=reduce)
-        else:
+        # if (
+        #     SYSTEM == "rocm"
+        #     and self.hidden_act == "silu"
+        #     and hidden_states.dtype == torch.float16
+        #     and hidden_states.shape[0] == 1
+        #     and not self.quantize
+        # ):
+        #     out = torch.empty(
+        #         hidden_states.shape[0],
+        #         self.intermediate_size,
+        #         dtype=hidden_states.dtype,
+        #         device="cuda",
+        #     )
+        #     _custom_C.LLMM_Silu(self.gate_up_proj.linear.weight, hidden_states, out, 8)
+        #     return self.down_proj(out, reduce=reduce)
+        # else:
             gate_up_states = self.gate_up_proj(hidden_states)
             gate_up_states = gate_up_states.view(-1, 2, self.intermediate_size)
             return self.down_proj(

server/text_generation_server/models/custom_modeling/flash_gptj_modeling.py

@@ -91,7 +91,7 @@ class GPTJRotary(PositionRotaryEmbedding):
 
             rotary_emb.apply_rotary(k1, k2, cos, sin, k1, k2, False)
         elif SYSTEM == "rocm":
-            from vllm._C import ops
+            import vllm._custom_ops as ops
 
             # NOTE: On RoCm systems, we use a ROPE implementatation adapted from VLLM which launches a single kernel for both query/key, contrary to flash-attn implementation used on NVIDIA systems.
             # Compiling flash-attn rotary on RoCm, it appears hipcc is unable to unroll loops, resulting in an even slower inference compared to eager: https://github.com/pytorch/pytorch/issues/113773
@@ -99,7 +99,7 @@ class GPTJRotary(PositionRotaryEmbedding):
             head_size = query.shape[-1]
 
             # Inplace operation, updating query and key.
-            ops.rotary_embedding(query, key, head_size, cos, sin, False)
+            torch.ops._C.rotary_embedding_tgi(query, key, head_size, cos, sin, True)
         elif SYSTEM == "ipex":
             import intel_extension_for_pytorch as ipex
 

server/text_generation_server/models/custom_modeling/flash_llama_modeling.py

@@ -64,7 +64,7 @@ if SYSTEM != "ipex":
 
 if SYSTEM == "rocm":
     try:
-        from vllm import _custom_C
+        from vllm import _custom_ops
     except Exception as e:
         raise ImportError(f"Could not load `vllm._custom_C`. Full error: {e}")
 
@@ -377,26 +377,26 @@ class LlamaMLP(nn.Module):
         self.hidden_size = config.hidden_size
 
     def forward(self, hidden_states, adapter_data):
-        if (
-            SYSTEM == "rocm"
-            and self.hidden_act == "silu"
-            and hidden_states.dtype == torch.float16
-            and hidden_states.shape[0] == 1
-            and not self.quantize
-            and self.hidden_size
-            != 16384  # TODO: Temporary workaround for `LLMM_Silu` kernel not working with LLama3.1 405B; needs refactoring once fixed.
-        ):
-            out = torch.empty(
-                hidden_states.shape[0],
-                self.intermediate_size,
-                dtype=hidden_states.dtype,
-                device="cuda",
-            )
-            _custom_C.LLMM_Silu(
-                self.gate_up_proj.base_layer.linear.weight, hidden_states, out, 8
-            )
-            return self.down_proj(out, adapter_data)
-        else:
+        # if (
+        #     SYSTEM == "rocm"
+        #     and self.hidden_act == "silu"
+        #     and hidden_states.dtype == torch.float16
+        #     and hidden_states.shape[0] == 1
+        #     and not self.quantize
+        #     and self.hidden_size
+        #     != 16384  # TODO: Temporary workaround for `LLMM_Silu` kernel not working with LLama3.1 405B; needs refactoring once fixed.
+        # ):
+        #     out = torch.empty(
+        #         hidden_states.shape[0],
+        #         self.intermediate_size,
+        #         dtype=hidden_states.dtype,
+        #         device="cuda",
+        #     )
+        #     _custom_C.LLMM_Silu(
+        #         self.gate_up_proj.base_layer.linear.weight, hidden_states, out, 8
+        #     )
+        #     return self.down_proj(out, adapter_data)
+        # else:
             gate_up_states = self.gate_up_proj(hidden_states, adapter_data)
             gate_up_states = gate_up_states.view(-1, 2, self.intermediate_size)
             return self.down_proj(

server/text_generation_server/models/custom_modeling/flash_mistral_modeling.py

@@ -49,7 +49,7 @@ from text_generation_server.layers.layernorm import (
 
 if SYSTEM == "rocm":
     try:
-        from vllm import _custom_C
+        from vllm import _custom_ops
     except Exception as e:
         raise ImportError(f"Could not load `vllm._custom_C`. Full error: {e}")
 
@@ -305,24 +305,24 @@ class MistralMLP(nn.Module):
         self.quantize = config.quantize
 
     def forward(self, hidden_states, adapter_data):
-        if (
-            SYSTEM == "rocm"
-            and self.hidden_act == "silu"
-            and hidden_states.dtype == torch.float16
-            and hidden_states.shape[0] == 1
-            and not self.quantize
-        ):
-            out = torch.empty(
-                hidden_states.shape[0],
-                self.intermediate_size,
-                dtype=hidden_states.dtype,
-                device="cuda",
-            )
-            _custom_C.LLMM_Silu(
-                self.gate_up_proj.base_layer.linear.weight, hidden_states, out, 8
-            )
-            return self.down_proj(out, adapter_data)
-        else:
+        # if (
+        #     SYSTEM == "rocm"
+        #     and self.hidden_act == "silu"
+        #     and hidden_states.dtype == torch.float16
+        #     and hidden_states.shape[0] == 1
+        #     and not self.quantize
+        # ):
+        #     out = torch.empty(
+        #         hidden_states.shape[0],
+        #         self.intermediate_size,
+        #         dtype=hidden_states.dtype,
+        #         device="cuda",
+        #     )
+        #     _custom_C.LLMM_Silu(
+        #         self.gate_up_proj.base_layer.linear.weight, hidden_states, out, 8
+        #     )
+        #     return self.down_proj(out, adapter_data)
+        # else:
             gate_up_states = self.gate_up_proj(hidden_states, adapter_data)
             gate_up_states = gate_up_states.view(-1, 2, self.intermediate_size)
             return self.down_proj(

server/text_generation_server/models/custom_modeling/idefics_modeling.py

@@ -52,7 +52,7 @@ from loguru import logger
 if SYSTEM == "cuda":
     import dropout_layer_norm
 elif SYSTEM == "rocm":
-    from vllm._C import ops
+    from vllm import _custom_ops
 else:
     dropout_layer_norm = None
 
@@ -424,7 +424,7 @@ class IdeficsRMSNorm(nn.Module):
                 hidden_states = hidden_states.reshape(-1, shape[-1])
 
             out = torch.empty_like(hidden_states)
-            ops.rms_norm(
+            _custom_ops.rms_norm(
                 out,
                 hidden_states,
                 self.weight.data,