xc-llm-ascend/vllm_ascend/platform.py

#
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
#
# 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.
# This file is a part of the vllm-ascend project.
#

from __future__ import annotations

import math
import os
from typing import TYPE_CHECKING, Any
from uuid import uuid4

import torch
import vllm.envs as envs_vllm
from vllm.logger import logger
from vllm.platforms import Platform, PlatformEnum

# todo: please remove it when solve cuda hard code in vllm
os.environ["VLLM_DISABLE_SHARED_EXPERTS_STREAM"] = "1"

from vllm_ascend.ascend_config import init_ascend_config

# isort: off
from vllm_ascend.utils import (
    ASCEND_QUANTIZATION_METHOD,
    COMPILATION_PASS_KEY,
    COMPRESSED_TENSORS_METHOD,
    AscendDeviceType,
    check_kv_extra_config,
    enable_sp,
    flashcomm2_enable,
    get_ascend_device_type,
    is_moe_model,
    is_vl_model,
    refresh_block_size,
    update_aclgraph_sizes,
    update_cudagraph_capture_sizes,
    update_default_aclgraph_sizes,
    is_310p,
)

if TYPE_CHECKING:
    from vllm.config import ModelConfig, VllmConfig
    from vllm.utils import FlexibleArgumentParser
else:
    ModelConfig = None
    VllmConfig = None
    FlexibleArgumentParser = None

_CUSTOM_OP_REGISTERED = False


def config_deprecated_logging():
    """Configure deprecated logging format, when used deprecated codes
    in vllm-ascend.
    """
    import logging
    import warnings

    # Customize warning format to be one line
    def one_line_formatwarning(message, category, filename, lineno, line=None):
        return f"{filename}:{lineno}: {category.__name__}: {message}"

    warnings.formatwarning = one_line_formatwarning

    logging.captureWarnings(True)
    warnings.simplefilter("once", DeprecationWarning)

    vllm_logger = logging.getLogger("vllm")
    warnings_logger = logging.getLogger("py.warnings")

    # Propagate vllm logger handlers to warnings logger, to keep the same
    # format with vllm
    if vllm_logger.handlers:
        warnings_logger.handlers = []

        for handler in vllm_logger.handlers:
            warnings_logger.addHandler(handler)

    warnings_logger.propagate = False


class NPUPlatform(Platform):
    _enum = PlatformEnum.OOT
    device_name: str = "npu"
    device_type: str = "npu"
    simple_compile_backend: str = "eager"  # Disable torch.compile()
    ray_device_key: str = "NPU"
    device_control_env_var: str = "ASCEND_RT_VISIBLE_DEVICES"
    dispatch_key: str = "PrivateUse1"

    supported_quantization: list[str] = [ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD]

    def is_sleep_mode_available(self) -> bool:
        return True

    @property
    def pass_key(self) -> str:
        """
        Inductor config key for the PassManager custom pass, for example 'post_grad_custom_post_pass'.
        It is a parameter of inductor_config used to register custom passes.
        Currently, we only use Inductor's 'pattern matcher' functionality, so we define our own pass_key.
        """
        return COMPILATION_PASS_KEY

    @classmethod
    def get_pass_manager_cls(cls) -> str:
        """
        Get the pass manager class for this platform.
        It will be registered as a custom pass under the current_platform.pass_key.
        """
        return "vllm_ascend.compilation.graph_fusion_pass_manager.GraphFusionPassManager"

    @classmethod
    def get_compile_backend(self) -> str:
        """
        Get the custom compile backend. Previously, we used EagerAdaptor by default.
        To use graph fusion operations, we defined our own backend compiler.
        """
        return "vllm_ascend.compilation.compiler_interface.AscendCompiler"

    @classmethod
    def pre_register_and_update(cls, parser: FlexibleArgumentParser | None = None) -> None:
        # Adapt the global patch here.
        from vllm_ascend.utils import adapt_patch

        adapt_patch(is_global_patch=True)

        # For online serving, "ascend" quantization method is not a choice natively,
        # so we need to add "ascend" quantization method to quantization methods list
        # and the user can enable quantization using "vllm serve --quantization ascend".
        if parser is not None:
            quant_action = parser._option_string_actions.get("--quantization")
            if quant_action and hasattr(quant_action, "choices") and quant_action.choices:
                if ASCEND_QUANTIZATION_METHOD not in quant_action.choices:
                    quant_action.choices.append(ASCEND_QUANTIZATION_METHOD)

        from vllm_ascend.quantization.compressed_tensors.compressed_tensors import AscendCompressedTensorsConfig  # noqa: F401
        from vllm_ascend.quantization.quant_config import AscendQuantConfig  # noqa: F401

        config_deprecated_logging()

    @classmethod
    def get_device_capability(cls, device_id: int = 0):
        return None

    @classmethod
    def get_device_name(cls, device_id: int = 0) -> str:
        return torch.npu.get_device_name(device_id)

    @classmethod
    def inference_mode(cls):
        return torch.inference_mode()

    @classmethod
    def set_device(cls, device: torch.device):
        torch.npu.set_device(device)

    @classmethod
    def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
        # initialize ascend config from vllm additional_config
        ascend_config = init_ascend_config(vllm_config)

        if vllm_config.kv_transfer_config is not None:
            check_kv_extra_config(vllm_config)
            if not getattr(vllm_config.kv_transfer_config, "_engine_id_patched", False):
                vllm_config.kv_transfer_config.engine_id = f"{vllm_config.kv_transfer_config.engine_id}-{uuid4().hex}"
                vllm_config.kv_transfer_config._engine_id_patched = True
        from vllm.config import CompilationMode  # noqa: E402

        compilation_config = vllm_config.compilation_config
        model_config = vllm_config.model_config
        parallel_config = vllm_config.parallel_config
        cache_config = vllm_config.cache_config
        ascend_compilation_config = ascend_config.ascend_compilation_config
        if ascend_compilation_config:
            vllm_config.additional_config.setdefault("ascend_compilation_config", {}).update(
                vars(ascend_compilation_config)
                if not isinstance(ascend_compilation_config, dict)
                else ascend_compilation_config
            )

        elif model_config and hasattr(model_config.hf_text_config, "index_topk"):
            vllm_config.cache_config.cache_dtype = str(model_config.dtype).replace("torch.", "")
        if model_config is None:
            logger.warning("Model config is missing. This may indicate that we are running a test case")
            enforce_eager = False
        else:
            enforce_eager = getattr(model_config, "enforce_eager", False)

        from vllm.config.compilation import CUDAGraphMode

        if enforce_eager:
            logger.info("Compilation disabled, using eager mode by default")
            compilation_config.mode = CompilationMode.NONE
            if compilation_config.splitting_ops is None:
                compilation_config.splitting_ops = []

        compilation_config.cudagraph_num_of_warmups = 1

        if compilation_config.mode not in [CompilationMode.NONE, CompilationMode.VLLM_COMPILE]:
            logger.warning(
                "NPU does not support %s compilation mode. Setting CUDAGraphMode to NONE", compilation_config.mode
            )
            compilation_config.cudagraph_mode = CUDAGraphMode.NONE

        # set cudaprah sizes before extending `compilation_config.splitting_ops`
        vllm_config._set_cudagraph_sizes()
        # There are cases where default cudagraph_capture_sizes are not friendly
        # to ascend ops && hardwares. We update these sizes here to improve
        # default performance.
        update_default_aclgraph_sizes(vllm_config)
        # TODO delete graph size update here when compilation_config.pass_config.enable_sp
        # is supported by vllm-ascend.
        if (
            vllm_config.parallel_config.tensor_parallel_size > 1
            and not vllm_config.model_config.enforce_eager
            and enable_sp(vllm_config)
        ):
            original_sizes = compilation_config.cudagraph_capture_sizes
            sp_aclgraph_sizes = vllm_config.update_sizes_for_sequence_parallelism(original_sizes)
            assert sp_aclgraph_sizes, (
                f"cudagraph_capture_sizes {original_sizes} does not contain"
                f"values that are multiples of tp_size "
                f"{vllm_config.parallel_config.tensor_parallel_size}"
            )
            if len(sp_aclgraph_sizes) != len(original_sizes):
                compilation_config.cudagraph_capture_sizes = sp_aclgraph_sizes
                update_cudagraph_capture_sizes(vllm_config, sp_aclgraph_sizes)

        # TODO: Full graph is fully supported later, and the default value will be set to full graph.
        if compilation_config.cudagraph_mode == CUDAGraphMode.FULL_AND_PIECEWISE:
            compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE

        # encoder-decoder models currently only support piecewise mode
        if model_config and model_config.is_encoder_decoder is True:
            if compilation_config.cudagraph_mode == CUDAGraphMode.FULL_DECODE_ONLY:
                logger.warning("encoder-decoder model doesn't support FULL_DECODE_ONLY, fallback to PIECEWISE ")
            compilation_config.cudagraph_mode = CUDAGraphMode.PIECEWISE

        # get custom compile backend for graph fusion
        compilation_config.oot_compiler = cls.get_compile_backend()

        if compilation_config.cudagraph_mode == CUDAGraphMode.NONE:
            compilation_config.mode = CompilationMode.NONE
            ascend_config.enable_npugraph_ex = False
        elif compilation_config.cudagraph_mode == CUDAGraphMode.PIECEWISE:
            logger.info("PIECEWISE compilation enabled on NPU. use_inductor not supported - using only ACL Graph mode")
            assert compilation_config.mode == CompilationMode.VLLM_COMPILE, (
                "When enabling VLLM_COMPILE aclgraph, please make sure compilation_config.mode == "
                "CompilationMode.VLLM_COMPILE and compilation_config.cudagraph_mode == CUDAGraphMode.VLLM_COMPILE"
            )
            compilation_config.set_splitting_ops_for_v1(
                all2all_backend=vllm_config.parallel_config.all2all_backend,
                data_parallel_size=vllm_config.parallel_config.data_parallel_size,
            )
            compilation_config.use_inductor = False
            # NOTE: Theoretically, we should also add vllm::mla_forward in the attention ops.
            # Since the process is created in the spawn mode, the value of the class attribute
            # attention ops transmitted is still the one before modification, so it has not been modified.
            # This will cause in scenarios where both piecewise and splitting ops are configured simultaneously,
            # If splitting ops does not contain the vllm::mla forward value, this configuration issue will
            # not be detected in advance assert.
            compilation_config.splitting_ops.extend(["vllm::mla_forward"])
            update_aclgraph_sizes(vllm_config)
            ascend_config.enable_npugraph_ex = False
        elif (
            compilation_config.cudagraph_mode == CUDAGraphMode.FULL_DECODE_ONLY
            or compilation_config.cudagraph_mode == CUDAGraphMode.FULL
        ):
            logger.info(
                "FULL_DECODE_ONLY compilation enabled on NPU. use_inductor not supported - using only ACL Graph mode"
            )
            compilation_config.use_inductor = False
            compilation_config.splitting_ops = []
            warning_message = """\033[91m
            **********************************************************************************
            * WARNING: You have enabled the *full graph* feature.
            * This is an early experimental stage and may involve various unknown issues.
            * A known problem is that capturing too many batch sizes can lead to OOM
            * (Out of Memory) errors or inference hangs. If you encounter such issues,
            * consider reducing `gpu_memory_utilization` or manually specifying a smaller
            * batch size for graph capture.
            * For more details, please refer to:
            * https://docs.vllm.ai/en/stable/configuration/conserving_memory.html#reduce-cuda-graphs
            **********************************************************************************\033[0m
            """
            logger.warning(warning_message)
        else:
            logger.info(
                "%s cudagraph_mode is not support on NPU. falling back to NONE", compilation_config.cudagraph_mode
            )
            compilation_config.cudagraph_mode = CUDAGraphMode.NONE
            compilation_config.mode = CompilationMode.NONE
            ascend_config.enable_npugraph_ex = False

        # TODO: Remove this check when ACL Graph supports ASCEND_LAUNCH_BLOCKING=1
        # Then, we will have to discuss the error handling strategy and user experience
        if (
            compilation_config.cudagraph_mode != CUDAGraphMode.NONE
            and os.environ.get("ASCEND_LAUNCH_BLOCKING", "0") == "1"
        ):
            raise ValueError(
                "ACL graph is incompatible with ASCEND_LAUNCH_BLOCKING=1. "
                "Please unset ASCEND_LAUNCH_BLOCKING or set it to 0. If you "
                "need ASCEND_LAUNCH_BLOCKING for debugging, consider other methods — "
                "for example, check the plog files (default: $HOME/ascend/log/debug) "
                "for more information about runtime errors."
            )

        if parallel_config and parallel_config.worker_cls == "auto":
            # TODO: this is a tricky way to disable `use_sequence_parallel_moe` in vllm.
            parallel_config.all2all_backend = "flashinfer_all2allv"
            if is_310p():
                parallel_config.worker_cls = "vllm_ascend._310p.worker_310p.NPUWorker310"
            elif ascend_config.xlite_graph_config.enabled:
                logger.info("openEuler Xlite enabled. See: https://atomgit.com/openeuler/GVirt/tree/master/xlite")
                parallel_config.worker_cls = "vllm_ascend.xlite.xlite_worker.XliteWorker"
            else:
                parallel_config.worker_cls = "vllm_ascend.worker.worker.NPUWorker"

        refresh_block_size(vllm_config)

        # Activate custom ops for v1, except on 310P
        if get_ascend_device_type() != AscendDeviceType._310P:
            compilation_config.custom_ops = ["all"]

        if ascend_config.recompute_scheduler_enable:
            from vllm_ascend.core.recompute_scheduler import RecomputeSchedulerConfig

            recompute_scheduler_config = RecomputeSchedulerConfig.initialize_from_config(vllm_config)
            vllm_config.scheduler_config = recompute_scheduler_config

        # Extend original scheduler_config to use SchedulerDynamicBatch.
        if ascend_config.SLO_limits_for_dynamic_batch != -1:
            vllm_config.scheduler_config.scheduler_cls = (
                "vllm_ascend.core.scheduler_dynamic_batch.SchedulerDynamicBatch"
            )
            vllm_config.scheduler_config.enable_chunked_prefill = True
            vllm_config.scheduler_config.SLO_limits_for_dynamic_batch = ascend_config.SLO_limits_for_dynamic_batch

        if (
            vllm_config.kv_transfer_config is not None
            and cache_config.block_size != parallel_config.cp_kv_cache_interleave_size
            and parallel_config.decode_context_parallel_size * parallel_config.prefill_context_parallel_size > 1
        ):
            raise AssertionError(
                f"cp_kv_cache_interleave_size({parallel_config.cp_kv_cache_interleave_size}) "
                f"and block_size({cache_config.block_size}) "
                "needs to be equal if use pcp or dcp > 1 in P/D disaggregate and kv pool scenario."
            )

        if is_vl_model(vllm_config):
            if bool(int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM", "0"))) or bool(
                int(os.getenv("VLLM_ASCEND_ENABLE_FLASHCOMM1", "0"))
            ):
                raise ValueError(
                    "Currently, VL models doesn't support "
                    "FLASHCOMM in vllm-ascend. We will fix this in the future. "
                    "Please set VLLM_ASCEND_ENABLE_FLASHCOMM1=0."
                )

    @classmethod
    def import_kernels(cls) -> None:
        # Directly importing vllm_ascend_C prevents ASCEND_RT_VISIBLE_DEVICES
        # from being applied during runtime initialization, which causes bugs
        # in the RL module. Therefore, we currently use lazy initialization
        # to avoid this issue. See https://github.com/vllm-project/vllm-ascend/pull/884.
        # TODO: when the above issue is fixed, we can uncomment the following lines.
        # from vllm_ascend.utils import enable_custom_op
        # enable_custom_op()
        # set custom ops path
        global _CUSTOM_OP_REGISTERED
        if _CUSTOM_OP_REGISTERED:
            return
        CUR_DIR = os.path.dirname(os.path.realpath(__file__))
        CUSTOM_OPP_PATH = os.path.join(CUR_DIR, "_cann_ops_custom", "vendors", "vllm-ascend")
        if os.path.exists(CUSTOM_OPP_PATH):
            current_cust_opp_path = os.environ.get("ASCEND_CUSTOM_OPP_PATH", "")
            if current_cust_opp_path:
                os.environ["ASCEND_CUSTOM_OPP_PATH"] = f"{CUSTOM_OPP_PATH}:{current_cust_opp_path}"
            else:
                os.environ["ASCEND_CUSTOM_OPP_PATH"] = CUSTOM_OPP_PATH
        _CUSTOM_OP_REGISTERED = True

    @classmethod
    def get_attn_backend_cls(cls, selected_backend, attn_selector_config):
        key = (attn_selector_config.use_mla, attn_selector_config.use_sparse)

        backend_map = {
            (True, False): "vllm_ascend.attention.mla_v1.AscendMLABackend",
            (False, False): "vllm_ascend.attention.attention_v1.AscendAttentionBackend",
            (True, True): "vllm_ascend.attention.sfa_v1.AscendSFABackend",
        }
        backend_map_310 = {
            (
                False,
                False,
            ): "vllm_ascend._310p.attention.attention_v1.AscendAttentionBackend310",
            # TODO If MLA/SFA is supported in the future, consider implementing the logic described in these comments.
            # (True, False): "...AscendMLABackend310",
            # (True, True):  "...AscendSFABackend310",
        }

        if is_310p():
            return backend_map_310.get(key, backend_map_310[(False, False)])

        return backend_map[key]

    @classmethod
    def get_punica_wrapper(cls) -> str:
        return "vllm_ascend.lora.punica_npu.PunicaWrapperNPU"

    @classmethod
    def get_current_memory_usage(cls, device: torch.types.Device | None = None) -> float:
        torch.npu.reset_peak_memory_stats(device)
        return torch.npu.max_memory_allocated(device)

    @classmethod
    def get_device_communicator_cls(cls) -> str:
        return "vllm_ascend.distributed.device_communicators.npu_communicator.NPUCommunicator"

    @classmethod
    def is_pin_memory_available(cls):
        return True

    @classmethod
    def opaque_attention_op(cls) -> bool:
        return True

    @classmethod
    def get_static_graph_wrapper_cls(cls) -> str:
        """
        Get piecewise backend class for piecewise graph.
        """
        return "vllm_ascend.compilation.acl_graph.ACLGraphWrapper"  # noqa

    @classmethod
    def support_hybrid_kv_cache(cls) -> bool:
        return True

    @classmethod
    def support_static_graph_mode(cls) -> bool:
        return True

    @classmethod
    def set_additional_forward_context(
        cls,
        attn_metadata: dict[str, Any],
        vllm_config: VllmConfig,
        dp_metadata,
        virtual_engine: int = 0,
        num_tokens: int | None = None,
        num_tokens_across_dp: torch.Tensor | None = None,
        cudagraph_runtime_mode=None,
        batch_descriptor=None,
        ubatch_slices=None,
    ) -> dict[str, Any]:
        """set additional forward context for ascend npus.

        Args:
            attn_metadata (dict[str, Any]): attention metadata for all layers.
            vllm_config (VllmConfig): configuration of vllm.
            dp_metadata (DpMetada): metadata for data parallelism.
                lack of typehint because of circular import.
            virtual_engine (int, optional): index of virtual engine. Defaults to 0.
            num_tokens (int | None, optional): number of tokens. Defaults to None.
            num_tokens_across_dp (torch.Tensor | None, optional): number of tokens
                across data parallelism.Defaults to None.
            cudagraph_runtime_mode (CUDAGraphMode, optional): mode of cudagraph runtime.
                Defaults to None.lack of typehint because of circular import.
            batch_descriptor (BatchDescriptor, optional): descriptor of batch.
                Defaults to None.
            ubatch_slices (UBatchSlices, optional): slice info for dual batch.
                Defaults to None. lack of typehint because of circular import

        Returns:
            dict[str, Any]: _description_
        """
        # NOTE(Ronald1995): avoid circular import.
        from vllm_ascend.ascend_forward_context import get_mc2_mask, select_moe_comm_method
        from vllm_ascend.ops.fused_moe.moe_comm_method import get_moe_comm_method
        from vllm.distributed import get_dp_group, get_tensor_model_parallel_world_size

        # NOTE(Ronald1995): avoid circular import, cudagraph_runtime_mode is
        # CUDAGraphMode.NONE in vllm, but we can't set CUDAGraphMode.NONE in
        # argument default value, so we set it to None first, then set it to
        # CUDAGraphMode.NONE here.
        from vllm.config import CUDAGraphMode

        if cudagraph_runtime_mode is None:
            cudagraph_runtime_mode = CUDAGraphMode.NONE
        # TODO(Ronald1995): model runner v1 still use ascend_forward_context,
        # when v1's forward context is refactored, we can remove this branch.
        # Currently, model runner v2 use the new forward context.
        # compared to v1, v2's forward context lacks some fields, such as:
        # in_profile_run, is_first_layer, prefetch_mlp_gate_up_proj,
        # prefetch_mlp_gate_down_proj, prefetch_mlp_enabled, model_instance,
        # is_draft_model.
        if not envs_vllm.VLLM_USE_V2_MODEL_RUNNER:
            return {}

        num_actual_tokens = list(attn_metadata.values())[0].num_actual_tokens
        if num_tokens is None:
            num_tokens = num_actual_tokens

        moe_comm_type = select_moe_comm_method(
            num_tokens,
            vllm_config,
            # is_draft_model will be removed later, so we set it to False temporarily.
            is_draft_model=False,
        )
        moe_comm_method = get_moe_comm_method(moe_comm_type)

        tp_world_size = get_tensor_model_parallel_world_size()

        # NOTE: This cannot be set using set_forward_context
        # due to multiple warmups before actual capturing.
        capturing = False

        # set for sequence parallelism, 1000 is the batch size concurrency
        # threshold for enabling the flashcomm_v1 or sequence_parallelism feature.
        # Currently, it is an empirical value. In normal scenarios,
        # if the concurrency exceeds this threshold,
        # the performance benefits can be maximized. Conversely,
        # if the concurrency is below the threshold,
        # the performance may degrade due to the switching of
        # communication methods.
        mmrs_fusion = True
        if is_moe_model(vllm_config):
            sp_enabled = enable_sp(vllm_config) and num_tokens is not None
            mmrs_fusion = False
        else:
            sp_enabled = enable_sp(vllm_config) and num_tokens is not None and num_tokens > 1000

        # TODO(Levi-JQ): another PR to normalize the enabling logic for sp/fc2
        flashcomm_v2_enabled = flashcomm2_enable() and tp_world_size > 1 and num_tokens is not None
        pad_size = 0
        if sp_enabled or flashcomm_v2_enabled:
            pad_size = (tp_world_size - (num_tokens % tp_world_size)) % tp_world_size

        dp_world_size = get_dp_group().world_size
        if dp_world_size > 1 and dp_metadata is not None:
            max_tokens_across_dp = dp_metadata.max_tokens_across_dp_cpu.item()
            if sp_enabled or flashcomm_v2_enabled:
                padded_length = (max_tokens_across_dp + tp_world_size - 1) // tp_world_size * tp_world_size
                pad_size = padded_length - num_tokens
        else:
            max_tokens_across_dp = num_tokens

        if num_tokens is not None:
            # NOTE: token num which need to pad to when mc2
            padded_num_tokens = math.ceil(max_tokens_across_dp / tp_world_size) * tp_world_size
            reserved_mc2_mask = get_mc2_mask()
            if reserved_mc2_mask is not None:
                mc2_mask = reserved_mc2_mask[:padded_num_tokens]
                mc2_mask[:num_actual_tokens] = True
                mc2_mask[num_actual_tokens:] = False
        return {
            "moe_comm_type": moe_comm_type,
            "moe_comm_method": moe_comm_method,
            "capturing": capturing,
            "mmrs_fusion": mmrs_fusion,
            "num_tokens": num_tokens,
            "sp_enabled": sp_enabled,
            "flashcomm_v2_enabled": flashcomm_v2_enabled,
            "pad_size": pad_size,
            "padded_length": padded_length,
            "max_tokens_across_dp": max_tokens_across_dp,
            "mc2_mask": mc2_mask,
        }