diff --git a/backends/exllamav3/model.py b/backends/exllamav3/model.py index 4f67fd3..53d2e91 100644 --- a/backends/exllamav3/model.py +++ b/backends/exllamav3/model.py @@ -25,6 +25,7 @@ from loguru import logger from backends.base_model_container import BaseModelContainer from backends.exllamav3.sampler import ExllamaV3SamplerBuilder +from backends.exllamav3.utils import exllama_supports_nccl from backends.exllamav3.vision import clear_image_embedding_cache from common.concurrency import iterate_in_threadpool from common.gen_logging import ( @@ -78,6 +79,7 @@ class ExllamaV3Container(BaseModelContainer): gpu_split_auto: bool = True autosplit_reserve: Optional[List[float]] = [96 / 1024] use_tp: bool = False + tp_backend: str = "native" max_seq_len: int = 4096 cache_size: int = 4096 cache_mode: str = "FP16" @@ -163,13 +165,25 @@ class ExllamaV3Container(BaseModelContainer): gpu_split_auto = unwrap(kwargs.get("gpu_split_auto"), True) gpu_split = unwrap(kwargs.get("gpu_split"), None) gpu_device_list = list(range(0, gpu_count)) + use_tp = unwrap(kwargs.get("tensor_parallel"), False) # Set GPU split options if gpu_count == 1: self.gpu_split_auto = False logger.info("Disabling GPU split because one GPU is in use.") else: - # TODO: Set tensor parallel + # Set tensor parallel + if use_tp: + self.use_tp = True + tp_backend = unwrap(kwargs.get("tensor_parallel_backend"), "native") + + if not exllama_supports_nccl(): + tp_backend = "native" + + self.tp_backend = tp_backend + + # TP has its own autosplit loader + self.gpu_split_auto = False # Set GPU split options # Enable manual GPU split if provided @@ -451,6 +465,8 @@ class ExllamaV3Container(BaseModelContainer): yield value for value in self.model.load_gen( + tensor_p=self.use_tp, + tp_backend=self.tp_backend, reserve_per_device=self.autosplit_reserve, use_per_device=self.gpu_split, callback=progress_callback, diff --git a/common/config_models.py b/common/config_models.py index b2b36e4..9898c04 100644 --- a/common/config_models.py +++ b/common/config_models.py @@ -183,11 +183,20 @@ class ModelConfig(BaseConfigModel): tensor_parallel: Optional[bool] = Field( False, description=( - "Load model with tensor parallelism.\n" + "Load model with tensor parallelism (default: False).\n" "Falls back to autosplit if GPU split isn't provided.\n" "This ignores the gpu_split_auto value." ), ) + tensor_parallel_backend: Optional[str] = Field( + "native", + description=( + "Sets a backend type for tensor parallelism. (default: native).\n" + "Options: native, nccl\n" + "Native is recommended for PCIe GPUs\n" + "NCCL is recommended for NVLink." + ), + ) gpu_split_auto: Optional[bool] = Field( True, description=( diff --git a/config_sample.yml b/config_sample.yml index 97cafa6..890596e 100644 --- a/config_sample.yml +++ b/config_sample.yml @@ -87,6 +87,12 @@ model: # This ignores the gpu_split_auto value. tensor_parallel: false + # Sets a backend type for tensor parallelism. (default: native). + # Options: native, nccl + # Native is recommended for PCIe GPUs + # NCCL is recommended for NVLink. + tensor_parallel_backend: native + # Automatically allocate resources to GPUs (default: True). # Not parsed for single GPU users. gpu_split_auto: true diff --git a/endpoints/core/types/model.py b/endpoints/core/types/model.py index 8e50622..1e84336 100644 --- a/endpoints/core/types/model.py +++ b/endpoints/core/types/model.py @@ -90,6 +90,7 @@ class ModelLoadRequest(BaseModel): examples=[4096], ) tensor_parallel: Optional[bool] = None + tensor_parallel_backend: Optional[str] = "native" gpu_split_auto: Optional[bool] = None autosplit_reserve: Optional[List[float]] = None gpu_split: Optional[List[float]] = Field( diff --git a/pyproject.toml b/pyproject.toml index 4874fd1..a45de2a 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -58,65 +58,65 @@ dev = [ ] cu121 = [ # Torch (Extra index URLs not support in pyproject.toml) - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp313-cp313-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp312-cp312-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp311-cp311-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", - "torch @ https://download.pytorch.org/whl/cu128/torch-2.7.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp313-cp313-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp312-cp312-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp311-cp311-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", + "torch @ https://download.pytorch.org/whl/cu128/torch-2.8.0%2Bcu128-cp310-cp310-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", # Exl2 - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.7.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+cu128.torch2.8.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", # Exl3 - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", - "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.5/exllamav3-0.0.5+cu128.torch2.7.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", + "exllamav3 @ https://github.com/turboderp-org/exllamav3/releases/download/v0.0.6/exllamav3-0.0.6+cu128.torch2.8.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", # Windows FA2 from https://github.com/kingbri1/flash-attention/releases - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", + "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu128torch2.8.0cxx11abiFALSE-cp313-cp313-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.13'", + "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu128torch2.8.0cxx11abiFALSE-cp312-cp312-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.12'", + "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu128torch2.8.0cxx11abiFALSE-cp311-cp311-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.11'", + "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu128torch2.8.0cxx11abiFALSE-cp310-cp310-win_amd64.whl ; platform_system == 'Windows' and python_version == '3.10'", # Linux FA2 from https://github.com/kingbri1/flash-attention/releases - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", - "flash_attn @ https://github.com/kingbri1/flash-attention/releases/download/v2.7.4.post1/flash_attn-2.7.4.post1+cu128torch2.7.0cxx11abiFALSE-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", + "flash_attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", + "flash_attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", + "flash_attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", + "flash_attn @ https://github.com/Dao-AILab/flash-attention/releases/download/v2.8.3/flash_attn-2.8.3+cu12torch2.8cxx11abiFALSE-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", ] amd = [ # Torch triton for ROCm - "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.3.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", - "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.3.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", - "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.3.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", - "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.3.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", + "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.4.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", + "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.4.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", + "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.4.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", + "pytorch_triton_rocm @ https://download.pytorch.org/whl/pytorch_triton_rocm-3.4.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", # Torch - "torch @ https://download.pytorch.org/whl/rocm6.3/torch-2.7.0%2Brocm6.3-cp313-cp313-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", - "torch @ https://download.pytorch.org/whl/rocm6.3/torch-2.7.0%2Brocm6.3-cp312-cp312-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", - "torch @ https://download.pytorch.org/whl/rocm6.3/torch-2.7.0%2Brocm6.3-cp311-cp311-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", - "torch @ https://download.pytorch.org/whl/rocm6.3/torch-2.7.0%2Brocm6.3-cp310-cp310-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", + "torch @ https://download.pytorch.org/whl/rocm6.4/torch-2.8.0%2Brocm6.4-cp313-cp313-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", + "torch @ https://download.pytorch.org/whl/rocm6.4/torch-2.8.0%2Brocm6.4-cp312-cp312-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", + "torch @ https://download.pytorch.org/whl/rocm6.4/torch-2.8.0%2Brocm6.4-cp311-cp311-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", + "torch @ https://download.pytorch.org/whl/rocm6.4/torch-2.8.0%2Brocm6.4-cp310-cp310-manylinux_2_28_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", # Exl2 - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.3.torch2.7.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.3.torch2.7.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.3.torch2.7.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", - "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.3.torch2.7.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.4.torch2.8.0-cp313-cp313-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.13'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.4.torch2.8.0-cp312-cp312-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.12'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.4.torch2.8.0-cp311-cp311-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.11'", + "exllamav2 @ https://github.com/turboderp-org/exllamav2/releases/download/v0.3.2/exllamav2-0.3.2+rocm6.4.torch2.8.0-cp310-cp310-linux_x86_64.whl ; platform_system == 'Linux' and platform_machine == 'x86_64' and python_version == '3.10'", ] # MARK: Ruff options