Files
xc-llm-ascend/vllm_ascend/ascend_forward_context.py

341 lines
14 KiB
Python
Raw Permalink Normal View History

import math
from contextlib import contextmanager
from enum import Enum
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
from typing import Any
import torch
import vllm.envs as envs_vllm
from vllm.config import CUDAGraphMode, VllmConfig
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
from vllm.distributed import get_dp_group, get_ep_group, get_tensor_model_parallel_world_size
from vllm.forward_context import BatchDescriptor, get_forward_context, set_forward_context
import vllm_ascend.envs as envs_ascend
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
from vllm_ascend.utils import (
AscendDeviceType,
enable_sp,
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
flashcomm2_enable,
get_ascend_device_type,
has_layer_idx,
is_drafter_moe_model,
is_moe_model,
speculative_enable_dispatch_gmm_combine_decode,
vllm_version_is,
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
)
class MoECommType(Enum):
ALLGATHER = 0
MC2 = 1
ALLTOALL = 2
FUSED_MC2 = 3
@contextmanager
def set_ascend_forward_context(
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
attn_metadata: Any,
vllm_config: VllmConfig,
virtual_engine: int = 0,
num_tokens: int = 0,
num_tokens_across_dp: torch.Tensor | None = None,
in_profile_run: bool = False,
num_actual_tokens: int | None = None,
aclgraph_runtime_mode: CUDAGraphMode = CUDAGraphMode.NONE,
batch_descriptor: BatchDescriptor | None = None,
model_instance: torch.nn.Module = None,
is_draft_model=False,
[Main2Main] Upgrade vllm commit to 0123 (#6169) ### What this PR does / why we need it? 1. ✅ Upgrade vllm commit to: 0115 (8471b27df97c3eb79f891802fc0e858f8f7ac6a0) Modify import paths due to the refactors: https://github.com/vllm-project/vllm/pull/32245 https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21034239336/job/60490156965?pr=5913 2. ✅Upgrade vllm commit to: 0119 (9a1f16da1e423ede2c2f52a9850cbfbb39cefe96) Fix `WorkerProc.__init__() missing 1 required positional argument: 'is_driver_worker'` due to https://github.com/vllm-project/vllm/pull/28506 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21156263050/job/60841668755?5569 3. ✅Upgrade vllm commit to: 0120(148117ea2e689cd43df4be6892671a17cdae5833) 1. Add `skip_compiled` param in `set_forward_context` due to https://github.com/vllm-project/vllm/pull/30385 2. Modify `tests/ut/spec_decode/test_eagle_proposer.py` due to https://github.com/vllm-project/vllm/pull/24322 change `self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens + max_batch_size` 3. Modify UT import paths due to the refactors:https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21204851770/job/60999046946 4. ✅Upgrade vllm commit to: 0121(f23fb5a7c1b61350c5c40ca1115d3bf8cf2b8cc9) 1. vLLM switched `uses_mrope` from target to draft model config, making `positions`/`mrope_positions` mutually exclusive, breaking vllm-ascend's direct self.positions access and tests missing `draft_model_config.uses_mrope`. https://github.com/vllm-project/vllm/pull/32048 2. Moved bs_to_padded_graph_size from CompilationConfig to CudagraphDispatcher due to the refactor https://github.com/vllm-project/vllm/pull/30143 3. Remove unused `maybe_setup_kv_connector` due to https://github.com/vllm-project/vllm/pull/32077 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21217728738/job/61043738834 6. ✅Upgrade vllm commit to: 0122(8ebf271bb6d1e7e9b1a55be73d755ef1a57dbbe5) Updating FusedMoEParallelConfig (added enable_eplb) and FusedMoEConfig due to https://github.com/vllm-project/vllm/pull/32414 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21249922546/job/61148613054 8. ✅Upgrade vllm commit to: 0123(dc917cceb877dfd13f98c538c4c96158047d98bd) Setting temperature=0.0 due to the removal of the default temperature value in https://github.com/vllm-project/vllm/pull/32723 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21280796875 ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.14.0 - vLLM main: https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60 --------- Signed-off-by: wjunLu <wjunlu217@gmail.com> Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com> Co-authored-by: wjunLu <wjunlu217@gmail.com>
2026-01-27 08:44:36 +08:00
skip_compiled: bool = False,
max_tokens_across_pcp: int = 0,
draft_attn_metadatas=None,
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
):
"""A context manager that stores the current forward context,
can be attention metadata, etc.
We add some additional param into forward_context.
"""
[Main2Main] Upgrade vllm commit to 0123 (#6169) ### What this PR does / why we need it? 1. ✅ Upgrade vllm commit to: 0115 (8471b27df97c3eb79f891802fc0e858f8f7ac6a0) Modify import paths due to the refactors: https://github.com/vllm-project/vllm/pull/32245 https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21034239336/job/60490156965?pr=5913 2. ✅Upgrade vllm commit to: 0119 (9a1f16da1e423ede2c2f52a9850cbfbb39cefe96) Fix `WorkerProc.__init__() missing 1 required positional argument: 'is_driver_worker'` due to https://github.com/vllm-project/vllm/pull/28506 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21156263050/job/60841668755?5569 3. ✅Upgrade vllm commit to: 0120(148117ea2e689cd43df4be6892671a17cdae5833) 1. Add `skip_compiled` param in `set_forward_context` due to https://github.com/vllm-project/vllm/pull/30385 2. Modify `tests/ut/spec_decode/test_eagle_proposer.py` due to https://github.com/vllm-project/vllm/pull/24322 change `self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens + max_batch_size` 3. Modify UT import paths due to the refactors:https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21204851770/job/60999046946 4. ✅Upgrade vllm commit to: 0121(f23fb5a7c1b61350c5c40ca1115d3bf8cf2b8cc9) 1. vLLM switched `uses_mrope` from target to draft model config, making `positions`/`mrope_positions` mutually exclusive, breaking vllm-ascend's direct self.positions access and tests missing `draft_model_config.uses_mrope`. https://github.com/vllm-project/vllm/pull/32048 2. Moved bs_to_padded_graph_size from CompilationConfig to CudagraphDispatcher due to the refactor https://github.com/vllm-project/vllm/pull/30143 3. Remove unused `maybe_setup_kv_connector` due to https://github.com/vllm-project/vllm/pull/32077 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21217728738/job/61043738834 6. ✅Upgrade vllm commit to: 0122(8ebf271bb6d1e7e9b1a55be73d755ef1a57dbbe5) Updating FusedMoEParallelConfig (added enable_eplb) and FusedMoEConfig due to https://github.com/vllm-project/vllm/pull/32414 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21249922546/job/61148613054 8. ✅Upgrade vllm commit to: 0123(dc917cceb877dfd13f98c538c4c96158047d98bd) Setting temperature=0.0 due to the removal of the default temperature value in https://github.com/vllm-project/vllm/pull/32723 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21280796875 ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.14.0 - vLLM main: https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60 --------- Signed-off-by: wjunLu <wjunlu217@gmail.com> Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com> Co-authored-by: wjunLu <wjunlu217@gmail.com>
2026-01-27 08:44:36 +08:00
forward_context_kwargs = {
"attn_metadata": attn_metadata,
"vllm_config": vllm_config,
"num_tokens": num_tokens,
"num_tokens_across_dp": num_tokens_across_dp,
"cudagraph_runtime_mode": aclgraph_runtime_mode,
"batch_descriptor": batch_descriptor,
[Main2Main][Deps][Misc] Upgrade vLLM to v0.15.0 (#6470) ### What this PR does / why we need it? This PR upgrades the vLLM dependency from `v0.14.1` to `v0.15.0`. This involves: - Updating the `VLLM_TAG` in all `Dockerfile`. - Updating the vLLM version in `docs/source/conf.py`. - Removing conditional code paths specific to `v0.14.1` across the codebase, which simplifies maintenance. - Fix `TypeError: MMEncoderAttention.__init__() got an unexpected keyword argument 'multimodal_config'` due to https://github.com/vllm-project/vllm/pull/31972. - Fix `_shared_experts: 'NoneType' object is not callable` due to https://github.com/vllm-project/vllm/pull/32082 by https://github.com/vllm-project/vllm-ascend/pull/6335. - Fix `ReshapeAndCacheOperation setup failed!` due to https://github.com/vllm-project/vllm/pull/25954 by overriding attention metadata slots. This upgrade is necessary to keep the project aligned with the latest features, bug fixes, and API changes in the vLLM project. ### Does this PR introduce _any_ user-facing change? No, this is an internal dependency update and does not introduce any user-facing changes. ### How was this patch tested? CI is expected to pass with these changes, ensuring that all existing tests are successful with the new vLLM version. - vLLM version: v0.14.1 - vLLM main: https://github.com/vllm-project/vllm/commit/dc917cceb877dfd13f98c538c4c96158047d98bd co-authored-by: shen-shanshan <467638484@qq.com> --------- Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>
2026-02-02 15:57:55 +08:00
"skip_compiled": skip_compiled,
[Main2Main] Upgrade vllm commit to 0123 (#6169) ### What this PR does / why we need it? 1. ✅ Upgrade vllm commit to: 0115 (8471b27df97c3eb79f891802fc0e858f8f7ac6a0) Modify import paths due to the refactors: https://github.com/vllm-project/vllm/pull/32245 https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21034239336/job/60490156965?pr=5913 2. ✅Upgrade vllm commit to: 0119 (9a1f16da1e423ede2c2f52a9850cbfbb39cefe96) Fix `WorkerProc.__init__() missing 1 required positional argument: 'is_driver_worker'` due to https://github.com/vllm-project/vllm/pull/28506 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21156263050/job/60841668755?5569 3. ✅Upgrade vllm commit to: 0120(148117ea2e689cd43df4be6892671a17cdae5833) 1. Add `skip_compiled` param in `set_forward_context` due to https://github.com/vllm-project/vllm/pull/30385 2. Modify `tests/ut/spec_decode/test_eagle_proposer.py` due to https://github.com/vllm-project/vllm/pull/24322 change `self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens + max_batch_size` 3. Modify UT import paths due to the refactors:https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21204851770/job/60999046946 4. ✅Upgrade vllm commit to: 0121(f23fb5a7c1b61350c5c40ca1115d3bf8cf2b8cc9) 1. vLLM switched `uses_mrope` from target to draft model config, making `positions`/`mrope_positions` mutually exclusive, breaking vllm-ascend's direct self.positions access and tests missing `draft_model_config.uses_mrope`. https://github.com/vllm-project/vllm/pull/32048 2. Moved bs_to_padded_graph_size from CompilationConfig to CudagraphDispatcher due to the refactor https://github.com/vllm-project/vllm/pull/30143 3. Remove unused `maybe_setup_kv_connector` due to https://github.com/vllm-project/vllm/pull/32077 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21217728738/job/61043738834 6. ✅Upgrade vllm commit to: 0122(8ebf271bb6d1e7e9b1a55be73d755ef1a57dbbe5) Updating FusedMoEParallelConfig (added enable_eplb) and FusedMoEConfig due to https://github.com/vllm-project/vllm/pull/32414 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21249922546/job/61148613054 8. ✅Upgrade vllm commit to: 0123(dc917cceb877dfd13f98c538c4c96158047d98bd) Setting temperature=0.0 due to the removal of the default temperature value in https://github.com/vllm-project/vllm/pull/32723 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21280796875 ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.14.0 - vLLM main: https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60 --------- Signed-off-by: wjunLu <wjunlu217@gmail.com> Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com> Co-authored-by: wjunLu <wjunlu217@gmail.com>
2026-01-27 08:44:36 +08:00
}
if vllm_version_is("0.18.0"):
forward_context_kwargs["virtual_engine"] = virtual_engine
[Main2Main] Upgrade vllm commit to 0123 (#6169) ### What this PR does / why we need it? 1. ✅ Upgrade vllm commit to: 0115 (8471b27df97c3eb79f891802fc0e858f8f7ac6a0) Modify import paths due to the refactors: https://github.com/vllm-project/vllm/pull/32245 https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21034239336/job/60490156965?pr=5913 2. ✅Upgrade vllm commit to: 0119 (9a1f16da1e423ede2c2f52a9850cbfbb39cefe96) Fix `WorkerProc.__init__() missing 1 required positional argument: 'is_driver_worker'` due to https://github.com/vllm-project/vllm/pull/28506 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21156263050/job/60841668755?5569 3. ✅Upgrade vllm commit to: 0120(148117ea2e689cd43df4be6892671a17cdae5833) 1. Add `skip_compiled` param in `set_forward_context` due to https://github.com/vllm-project/vllm/pull/30385 2. Modify `tests/ut/spec_decode/test_eagle_proposer.py` due to https://github.com/vllm-project/vllm/pull/24322 change `self.max_num_tokens = vllm_config.scheduler_config.max_num_batched_tokens + max_batch_size` 3. Modify UT import paths due to the refactors:https://github.com/vllm-project/vllm/pull/32060 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21204851770/job/60999046946 4. ✅Upgrade vllm commit to: 0121(f23fb5a7c1b61350c5c40ca1115d3bf8cf2b8cc9) 1. vLLM switched `uses_mrope` from target to draft model config, making `positions`/`mrope_positions` mutually exclusive, breaking vllm-ascend's direct self.positions access and tests missing `draft_model_config.uses_mrope`. https://github.com/vllm-project/vllm/pull/32048 2. Moved bs_to_padded_graph_size from CompilationConfig to CudagraphDispatcher due to the refactor https://github.com/vllm-project/vllm/pull/30143 3. Remove unused `maybe_setup_kv_connector` due to https://github.com/vllm-project/vllm/pull/32077 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21217728738/job/61043738834 6. ✅Upgrade vllm commit to: 0122(8ebf271bb6d1e7e9b1a55be73d755ef1a57dbbe5) Updating FusedMoEParallelConfig (added enable_eplb) and FusedMoEConfig due to https://github.com/vllm-project/vllm/pull/32414 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21249922546/job/61148613054 8. ✅Upgrade vllm commit to: 0123(dc917cceb877dfd13f98c538c4c96158047d98bd) Setting temperature=0.0 due to the removal of the default temperature value in https://github.com/vllm-project/vllm/pull/32723 Test result: https://github.com/vllm-project/vllm-ascend/actions/runs/21280796875 ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.14.0 - vLLM main: https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60 --------- Signed-off-by: wjunLu <wjunlu217@gmail.com> Signed-off-by: Meihan-chen <jcccx.cmh@gmail.com> Co-authored-by: wjunLu <wjunlu217@gmail.com>
2026-01-27 08:44:36 +08:00
with set_forward_context(**forward_context_kwargs):
forward_context = get_forward_context()
forward_context.draft_attn_metadatas = draft_attn_metadatas
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
from vllm_ascend.ops.fused_moe.moe_comm_method import get_moe_comm_method
[Bugfix][eager][oom] fix rank0 load imbalance by no padding when multi dp (#7297) ### What this PR does / why we need it? Fix multi dp padding logic for eager mode, bacause its will cause rank0 load imbalance in kimi-k2.5-w4a8 with the all the padding tokens router to rank0. And the fix can also apply to other model in multi dp. - before hbm usage: <img width="2229" height="733" alt="image" src="https://github.com/user-attachments/assets/50479b6d-cfd0-4206-8e80-974024652997" /> preformance: ```shell Concurrency NumPrompts QPS TTFT_Avg TTFT_P50 TPOT_Avg TPOT_P50 TPOT_P90 ============ ============ ============ ============ ============ ============ ============ ============ 1 15 0.0179 1667.7803 1673.3437 35.2973 35.2775 35.3784 32 480 0.4725 2764.8027 1905.2137 40.8030 40.6978 41.0179 64 960 0.7820 4123.7096 3485.6153 48.0461 48.1598 48.2971 100 1500 1.0852 6216.7988 5714.0082 52.9323 53.0613 54.6304 108 1620 1.1040 6277.4892 5798.7425 56.3862 56.9224 57.2901 116 1740 1.1680 6563.3293 6039.5659 56.9894 57.4027 57.5786 128 1920 1.2555 7822.5551 7604.1662 57.7660 58.1768 58.2717 192 2880 1.4314 9212.1953 9131.3461 58.9905 59.1683 59.2791 256 3840 1.4480 9028.0812 8913.7937 59.0092 59.2385 59.3516 ``` - after hbm usage: <img width="2246" height="1005" alt="image" src="https://github.com/user-attachments/assets/d0936481-5a58-4bc5-a6f1-b92735d47885" /> preformance: ```shell Concurrency NumPrompts QPS TTFT_Avg TTFT_P50 TPOT_Avg TPOT_P50 TPOT_P90 ============ ============ ============ ============ ============ ============ ============ ============ 1 15 0.0181 601.4171 600.9774 35.6270 35.6254 35.6480 32 480 0.4455 720.8782 724.2889 45.4250 45.4755 45.6318 64 960 0.8445 729.6209 728.2149 47.0464 47.0896 47.1985 100 1500 1.2601 723.4834 724.6673 48.3108 48.3844 48.5355 108 1620 1.3409 727.1509 720.6772 48.8962 48.9409 49.0489 116 1740 1.4080 679.9799 677.6119 49.1253 49.1983 49.3087 128 1920 1.4155 680.6284 674.9436 49.2193 49.2450 49.3763 192 2880 1.4422 684.6577 676.7833 49.2059 49.2264 49.3229 256 3840 1.4558 685.2462 678.1709 49.2191 49.2351 49.3419 ``` ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.17.0 - vLLM main: https://github.com/vllm-project/vllm/commit/4034c3d32e30d01639459edd3ab486f56993876d --------- Signed-off-by: Levi-JQ <yujinqi2@huawei.com> Co-authored-by: Levi-JQ <yujinqi2@huawei.com> Co-authored-by: fny-coder <985619145@qq.com>
2026-03-23 17:05:02 +08:00
max_num_tokens = int(num_tokens_across_dp.max().item()) if num_tokens_across_dp is not None else num_tokens
moe_comm_type = select_moe_comm_method(max_num_tokens, vllm_config, is_draft_model)
forward_context.moe_comm_type = moe_comm_type
forward_context.moe_comm_method = get_moe_comm_method(moe_comm_type)
tp_world_size = get_tensor_model_parallel_world_size()
forward_context.in_profile_run = in_profile_run
# NOTE: This cannot be set using set_forward_context
# due to multiple warmups before actual capturing
forward_context.capturing = False
# TODO: remove it when torch_npu.npu_mm_reduce_scatter_base supports tp_size >= 16.
mmrs_fusion = tp_world_size <= 8
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
# 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.
# main model and drafter model may have different architecture
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
is_context_moe_model = is_drafter_moe_model(vllm_config) if is_draft_model else is_moe_model(vllm_config)
if is_context_moe_model:
flash_comm_v1_enabled = enable_sp(vllm_config) and num_tokens is not None
mmrs_fusion = False
[Bugfix] Avoided a bug of drafter when `dp` and `sp` are enabled (#6226) ### What this PR does / why we need it? Avoided a bug of drafter when `dp` and `sp` are enabled. Specifically, disable `sp` when drafter is dense. ### Does this PR introduce _any_ user-facing change? N/A ### How was this patch tested? An aisbench test: ```shell python3 aisbench_test.py --input_len 3500 --output_len 1000 --data_num 100 --concurrency 320 --request_rate 8 ``` The result is okay. ```text [2026-01-24 22:38:20,256] [ais_bench.benchmark.openicl.icl_inferencer.icl_gen_inferencer] [INFO] Calculate global interval offsets time: 0.5922 s 01/24 22:38:20 - AISBench - INFO - Process 0 using precomputed sleep offsets with 100 requests Process-0 pid:220279: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [09:40<00:00, 5.81s/it] Pid: 220279 | Post: 100 | Received: 100 | Failed: 0 | Post Time:12.51s | Receive Time:580.92s: Encoding output text...: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:01<00:00, 93.75it/s] 01/24 22:48:02 - AISBench - INFO - Start converting origin data to detailed data ... 01/24 22:48:02 - AISBench - INFO - Finish converting origin data to detailed data█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 100/100 [00:01<00:00, 95.08it/s] 01/24 22:48:02 - AISBench - INFO - Added 'Actual RPS: After Excluding Anomalies' to group 'Time - RPS: ' in legend explanation table 01/24 22:48:02 - AISBench - INFO - Successfully merged chart into position (1, 1) 01/24 22:48:02 - AISBench - INFO - RPS distribution charts saved to outputs/default/20260124_223809/performances/vllm-api-stream-chat/gsm8kdataset_rps_distribution_plot_with_actual_rps.html 01/24 22:48:02 - AISBench - INFO - Updated chart with actual RPS saved to outputs/default/20260124_223809/performances/vllm-api-stream-chat/gsm8kdataset_rps_distribution_plot_with_actual_rps.html [2026-01-24 22:48:02,557] [ais_bench.benchmark.openicl.icl_inferencer.icl_gen_perf_inferencer] [INFO] Start extracting pref datas ... [2026-01-24 22:48:02,558] [ais_bench.benchmark.openicl.icl_inferencer.icl_gen_perf_inferencer] [INFO] Finish extracting pref datas! [2026-01-24 22:48:02,558] [ais_bench.benchmark.openicl.icl_inferencer.icl_gen_perf_inferencer] [INFO] Dumping detail perf data ... Dumping data to h5: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 75.31it/s] [2026-01-24 22:48:02,588] [ais_bench.benchmark.openicl.icl_inferencer.icl_gen_perf_inferencer] [INFO] Dump detail perf data cost: 0.02995561994612217(s) [2026-01-24 22:48:02,588] [ais_bench.benchmark.openicl.icl_inferencer.icl_gen_perf_inferencer] [INFO] Performance task finished, results saved in outputs/default/20260124_223809/performances/vllm-api-stream-chat 01/24 22:48:02 - AISBench - INFO - time elapsed: 586.32s Running tasks: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [09:55<00:00, 595.91s/it] 01/24 22:48:05 - AISBench - INFO - Performance evaluation tasks completed. 01/24 22:48:05 - AISBench - INFO - Loading detail perf data of model='vllm-api-stream-chat' dataset='gsm8kdataset' ... 01/24 22:48:05 - AISBench - INFO - Starting request timeline processing... 01/24 22:48:05 - AISBench - INFO - Data preprocessing completed in 0.0004s 01/24 22:48:05 - AISBench - INFO - Generating timeline traces for 100 requests... 01/24 22:48:05 - AISBench - INFO - Generated timeline trace chunks in 0.0441s 01/24 22:48:05 - AISBench - INFO - Generating concurrency traces... 01/24 22:48:05 - AISBench - INFO - Generated concurrency trace chunks in 0.0011s 01/24 22:48:05 - AISBench - INFO - Creating figure layout... 01/24 22:48:05 - AISBench - INFO - Figure layout created in 0.0504s 01/24 22:48:05 - AISBench - INFO - Writing to outputs/default/20260124_223809/performances/vllm-api-stream-chat/gsm8kdataset_plot.html... 01/24 22:48:05 - AISBench - INFO - HTML written in 0.0181s 01/24 22:48:05 - AISBench - INFO - Completed! Total execution time: 0.1148s 01/24 22:48:05 - AISBench - INFO - The gsm8kdataset_plot has been saved in outputs/default/20260124_223809/performances/vllm-api-stream-chat/gsm8kdataset_plot.html 01/24 22:48:05 - AISBench - INFO - Converting perf results of stage ... 01/24 22:48:05 - AISBench - INFO - Finish Converting! 01/24 22:48:05 - AISBench - INFO - Start calculating metrics ... 01/24 22:48:05 - AISBench - INFO - Start calculating common metrics ... 01/24 22:48:05 - AISBench - INFO - Start calculating add units ... 01/24 22:48:05 - AISBench - INFO - Finish calculating perf data! 01/24 22:48:05 - AISBench - INFO - Summarizing performance results... 01/24 22:48:05 - AISBench - INFO - Performance Results of task: vllm-api-stream-chat/gsm8kdataset: ╒══════════════════════════╤═════════╤════════════════╤════════════════╤════════════════╤════════════════╤════════════════╤════════════════╤════════════════╤═════╕ │ Performance Parameters │ Stage │ Average │ Min │ Max │ Median │ P75 │ P90 │ P99 │ N │ ╞══════════════════════════╪═════════╪════════════════╪════════════════╪════════════════╪════════════════╪════════════════╪════════════════╪════════════════╪═════╡ │ E2EL │ total │ 300806.1781 ms │ 189326.0489 ms │ 568345.5121 ms │ 380629.6785 ms │ 384208.3527 ms │ 385363.7709 ms │ 566871.7684 ms │ 100 │ ├──────────────────────────┼─────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼─────┤ │ TTFT │ total │ 107441.2231 ms │ 343.8054 ms │ 378132.3979 ms │ 188817.4877 ms │ 190985.8451 ms │ 192547.6847 ms │ 378008.356 ms │ 100 │ ├──────────────────────────┼─────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼─────┤ │ TPOT │ total │ 193.5585 ms │ 185.1008 ms │ 197.262 ms │ 193.8146 ms │ 195.0803 ms │ 196.0323 ms │ 196.9688 ms │ 100 │ ├──────────────────────────┼─────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼─────┤ │ ITL │ total │ 194.2067 ms │ 0.0108 ms │ 2782.7124 ms │ 184.9998 ms │ 194.2631 ms │ 221.2895 ms │ 304.363 ms │ 100 │ ├──────────────────────────┼─────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼─────┤ │ InputTokens │ total │ 3506.86 │ 3431.0 │ 3508.0 │ 3508.0 │ 3508.0 │ 3508.0 │ 3508.0 │ 100 │ ├──────────────────────────┼─────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼─────┤ │ OutputTokens │ total │ 1000.0 │ 1000.0 │ 1000.0 │ 1000.0 │ 1000.0 │ 1000.0 │ 1000.0 │ 100 │ ├──────────────────────────┼─────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼─────┤ │ OutputTokenThroughput │ total │ 3.7745 token/s │ 1.7595 token/s │ 5.2819 token/s │ 2.6272 token/s │ 5.1028 token/s │ 5.1502 token/s │ 5.2754 token/s │ 100 │ ╘══════════════════════════╧═════════╧════════════════╧════════════════╧════════════════╧════════════════╧════════════════╧════════════════╧════════════════╧═════╛ ╒══════════════════════════╤═════════╤══════════════════╕ │ Common Metric │ Stage │ Value │ ╞══════════════════════════╪═════════╪══════════════════╡ │ Benchmark Duration │ total │ 580456.2704 ms │ ├──────────────────────────┼─────────┼──────────────────┤ │ Total Requests │ total │ 100 │ ├──────────────────────────┼─────────┼──────────────────┤ │ Failed Requests │ total │ 0 │ ├──────────────────────────┼─────────┼──────────────────┤ │ Success Requests │ total │ 100 │ ├──────────────────────────┼─────────┼──────────────────┤ │ Concurrency │ total │ 51.8224 │ ├──────────────────────────┼─────────┼──────────────────┤ │ Max Concurrency │ total │ 320 │ ├──────────────────────────┼─────────┼──────────────────┤ │ Request Throughput │ total │ 0.1723 req/s │ ├──────────────────────────┼─────────┼──────────────────┤ │ Total Input Tokens │ total │ 350686 │ ├──────────────────────────┼─────────┼──────────────────┤ │ Prefill Token Throughput │ total │ 32.6398 token/s │ ├──────────────────────────┼─────────┼──────────────────┤ │ Total generated tokens │ total │ 100000 │ ├──────────────────────────┼─────────┼──────────────────┤ │ Input Token Throughput │ total │ 604.1558 token/s │ ├──────────────────────────┼─────────┼──────────────────┤ │ Output Token Throughput │ total │ 172.2783 token/s │ ├──────────────────────────┼─────────┼──────────────────┤ │ Total Token Throughput │ total │ 776.434 token/s │ ╘══════════════════════════╧═════════╧══════════════════╛ 01/24 22:48:05 - AISBench - INFO - Performance Result files locate in outputs/default/20260124_223809/performances/vllm-api-stream-chat. ``` - vLLM version: v0.14.0 - vLLM main: https://github.com/vllm-project/vllm/commit/d68209402ddab3f54a09bc1f4de9a9495a283b60 Signed-off-by: drslark <slarksblood@qq.com>
2026-01-25 17:45:29 +08:00
elif is_draft_model:
# TODO: for dense drafter, `sp` is redundant and is not compatible with `dp` and `graph`.
# Disable it to avoid more problems.
flash_comm_v1_enabled = False
else:
flash_comm_v1_enabled = enable_sp(vllm_config) and num_tokens is not None and num_tokens > 1000
forward_context.mmrs_fusion = mmrs_fusion
forward_context.num_tokens = num_tokens
forward_context.flash_comm_v1_enabled = flash_comm_v1_enabled
# TODO(Levi-JQ): another PR to normalize the enabling logic for sp/fc2
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
forward_context.flashcomm_v2_enabled = flashcomm2_enable() and tp_world_size > 1 and num_tokens is not None
forward_context.pad_size = 0
if forward_context.flash_comm_v1_enabled or forward_context.flashcomm_v2_enabled:
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
pad_size = (tp_world_size - (num_tokens % tp_world_size)) % tp_world_size
forward_context.pad_size = pad_size
# set this for rope forward_oot using
forward_context.is_first_layer = True
# set layer_idx to enable optimization features that depend on this information.
# This is only applicable to models that contain these necessary attributes.
forward_context.layer_idx = None
if has_layer_idx(model_instance):
forward_context.layer_idx = model_instance.model.start_layer
forward_context.prefetch_mlp_gate_up_proj = False
forward_context.prefetch_mlp_down_proj = False
forward_context.model_instance = model_instance
forward_context.is_draft_model = is_draft_model
if num_tokens is None and attn_metadata is not None:
num_tokens = attn_metadata.num_actual_tokens
dp_world_size = get_dp_group().world_size
if dp_world_size > 1 and forward_context.dp_metadata is not None:
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
max_tokens_across_dp = forward_context.dp_metadata.max_tokens_across_dp_cpu.item()
if forward_context.flash_comm_v1_enabled or forward_context.flashcomm_v2_enabled:
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
padded_length = (max_tokens_across_dp + tp_world_size - 1) // tp_world_size * tp_world_size
pad_size = padded_length - num_tokens
forward_context.padded_length = padded_length
forward_context.pad_size = pad_size
else:
max_tokens_across_dp = num_tokens
forward_context.max_tokens_across_dp = max_tokens_across_dp
forward_context.max_tokens_across_pcp = max_tokens_across_pcp
if num_tokens is not None:
if num_actual_tokens is None:
num_actual_tokens = num_tokens
# NOTE: token num which need to pad to when mc2
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
forward_context.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:
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
mc2_mask = reserved_mc2_mask[: forward_context.padded_num_tokens]
mc2_mask[:num_actual_tokens] = True
mc2_mask[num_actual_tokens:] = False
forward_context.mc2_mask = mc2_mask
try:
yield
finally:
pass
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
_mc2_tokens_capacity: int | None = None
_reserved_mc2_mask: torch.Tensor | None = None
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
def set_mc2_tokens_capacity(vllm_config, max_num_reqs, uniform_decode_query_len):
global _mc2_tokens_capacity
if _mc2_tokens_capacity is not None:
return
if vllm_config.compilation_config.cudagraph_capture_sizes:
max_num_tokens = vllm_config.compilation_config.max_cudagraph_capture_size
else:
# NOTE: To save memory, we cap the max number of tokens to 512.
max_num_tokens = min(max_num_reqs * uniform_decode_query_len, 512)
tp_size = vllm_config.parallel_config.tensor_parallel_size
# Use integer arithmetic for ceiling division.
num_tokens_per_tp_rank = (max_num_tokens + tp_size - 1) // tp_size
_mc2_tokens_capacity = num_tokens_per_tp_rank * tp_size
def get_mc2_tokens_capacity():
return _mc2_tokens_capacity
def set_mc2_mask(vllm_config, device):
global _reserved_mc2_mask
if _reserved_mc2_mask is not None:
return
if is_moe_model(vllm_config):
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
_reserved_mc2_mask = torch.zeros(get_mc2_tokens_capacity(), dtype=torch.bool, device=device)
else:
_reserved_mc2_mask = None
def get_mc2_mask():
return _reserved_mc2_mask
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
def select_moe_comm_method(num_tokens: int, vllm_config: VllmConfig, is_draft_model=False) -> MoECommType | None:
"""Select the MoE communication method according to parallel settings,
device generation, token count, and quantization.
1. Non-MoE models return `None`.
2. Without expert parallel, fall back to all-gather.
3. On A2 with expert parallel, pick MC2 when tokens fit the MC2 capacity
and the DP size is large enough; otherwise use all-gather.
4. On A3 with expert parallel, prefer fused MC2 when using w8a8_dynamic
quantization with small EP size, no dynamic_eplb, and not in MTP
mode; otherwise use MC2 within capacity or all-to-all.
[Feat.]: 310p support MOE models (#6530) ### What this PR does / why we need it? This pull request integrates comprehensive support for Mixture of Experts (MoE) models on the Ascend 310P device within the vllm-ascend framework. It achieves this by introducing specialized modules for expert selection, fused MoE layers, and optimized all-gather communication. The changes also refine existing NPU operations, making them more consistent and efficient for 310P, ultimately enhancing the performance and compatibility of MoE models on this hardware. Highlights 310P MoE Support: Introduces dedicated implementations for Mixture of Experts (MoE) models on Ascend 310P devices, including new modules for expert selection, fused MoE layers, and communication. All-Gather Communication: Enforces the use of ALLGATHER communication for MoE operations on 310P, optimizing data transfer and leveraging NPU-specific token dispatching. Simplified NPU Operations: Removes conditional type casting for npu_swiglu and enables custom rotary embedding kernels unconditionally, suggesting improved native support for 310P. New MoE Classes Registered: Registers AscendFusedMoE310 and AscendSharedFusedMoE310 to integrate 310P-specific MoE layers into the system's custom operation registry. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? offline test and server test, with qwen3-30b-a3b,tp/ep 4 on 310p - vLLM version: v0.15.0 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0 --------- Signed-off-by: pu-zhe <zpuaa@outlook.com>
2026-02-06 10:30:56 +08:00
5. On 310P, always use all-gather.
Args:
num_tokens (int): The number of tokens in the current batch.
vllm_config (VllmConfig): Runtime configuration for the model.
is_draft_model (bool): Whether the model runs in MTP mode (disables fused MC2).
Raises:
ValueError: If the soc version is unsupported.
Returns:
MoECommType | None: The selected MoE communication method.
"""
if not is_moe_model(vllm_config):
return None
mc2_tokens_capacity = get_mc2_tokens_capacity()
soc_version = get_ascend_device_type()
quant_type = getattr(
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
vllm_config.model_config.hf_text_config,
"moe_quantize",
getattr(vllm_config.model_config.hf_text_config, "quantize", None),
)
[Lint]Style: Convert `vllm-ascend/compilation` to ruff format (#5912) ### What this PR does / why we need it? Convert `vllm-ascend/compilation` to ruff format. ### Does this PR introduce _any_ user-facing change? During this migration, we encountered some **errors** in our CI and testing environments, such as: ``` vllm_ascend/utils.py:653: in <module> def register_ascend_customop(vllm_config: VllmConfig | None = None): ^^^^^^^^^^^^^^^^^ E TypeError: unsupported operand type(s) for |: 'NoneType' and 'NoneType' ``` **1. Root Cause Analysis:** The project uses a common pattern to break circular dependencies: ```python if TYPE_CHECKING: from vllm.config import VllmConfig else: VllmConfig = None # Placeholder assigned at runtime ``` When Python parses the function definition `def register_ascend_customop(vllm_config: VllmConfig | None)`, it attempts to evaluate the expression `VllmConfig | None`. Since `VllmConfig` is assigned `None` at runtime, the expression effectively becomes `None | None`. In Python, `None` is an instance of `NoneType`. While the `|` operator is implemented for Type objects (classes), it is not supported for `NoneType` instances, leading to the `TypeError` shown above. **2. Solution:** To maintain the modern `|` syntax required by our new linting standards while preserving our dependency management strategy, I have introduced: ```python from __future__ import annotations ``` at the top of the affected files. This enables **Postponed Evaluation of Annotations (PEP 563)**. **3. Impact and Benefits:** - By enabling `annotations`, Python no longer executes the `VllmConfig | None` operation during module load. Instead, it stores the annotation as a string literal, completely avoiding the `None | None` calculation. - We can keep the `VllmConfig = None` placeholders. This ensures that other modules can still import these symbols without triggering an `ImportError`, maintaining a stable dependency graph. - IDEs and static type checkers (MyPy/Pyright) continue to resolve the types correctly. This allows us to use modern syntax without sacrificing type safety or runtime stability. - The only side effect is that `__annotations__` will now return strings instead of type objects. Since this module does not use runtime type enforcement or reflection, this change has zero negative impact on existing functionality. ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: https://github.com/vllm-project/vllm/commit/11b6af5280d6d6dfb8953af16e67b25f819b3be9 --------- Signed-off-by: MrZ20 <2609716663@qq.com>
2026-01-16 20:57:46 +08:00
if not vllm_config.parallel_config.enable_expert_parallel or get_ep_group().world_size == 1:
moe_comm_type = MoECommType.ALLGATHER
elif soc_version in {AscendDeviceType.A2}:
num_experts = vllm_config.model_config.get_num_experts()
ep_world_size = (
vllm_config.parallel_config.world_size_across_dp // vllm_config.parallel_config.pipeline_parallel_size
)
num_experts_per_device = num_experts // ep_world_size
if num_experts_per_device <= 24 and ep_world_size >= 16 and num_tokens <= mc2_tokens_capacity:
moe_comm_type = MoECommType.MC2
else:
moe_comm_type = MoECommType.ALLGATHER
elif soc_version in {AscendDeviceType.A3}:
# TODO: drop the EP-size guard when dispatch_ffn_combine supports larger EP sizes
# TODO: drop speculative method guard when dispatch_gmm_combine_decode supports w16a16
fused_mc2_enable = envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2
dispatch_ffn_combine_enable = get_ep_group().world_size <= 32 and (not is_draft_model)
if num_tokens <= mc2_tokens_capacity:
fused_decode_enable = fused_mc2_enable
if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 1:
fused_decode_enable = fused_mc2_enable and dispatch_ffn_combine_enable
elif envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 2:
fused_decode_enable = (
fused_mc2_enable
and speculative_enable_dispatch_gmm_combine_decode(vllm_config)
and quant_type == "w8a8_dynamic"
)
moe_comm_type = MoECommType.FUSED_MC2 if fused_decode_enable else MoECommType.MC2
else:
fused_prefill_enable = fused_mc2_enable
if envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 1:
fused_prefill_enable = fused_mc2_enable and dispatch_ffn_combine_enable
elif envs_ascend.VLLM_ASCEND_ENABLE_FUSED_MC2 == 2:
fused_prefill_enable = False
moe_comm_type = MoECommType.FUSED_MC2 if fused_prefill_enable else MoECommType.ALLTOALL
[Feat.]: 310p support MOE models (#6530) ### What this PR does / why we need it? This pull request integrates comprehensive support for Mixture of Experts (MoE) models on the Ascend 310P device within the vllm-ascend framework. It achieves this by introducing specialized modules for expert selection, fused MoE layers, and optimized all-gather communication. The changes also refine existing NPU operations, making them more consistent and efficient for 310P, ultimately enhancing the performance and compatibility of MoE models on this hardware. Highlights 310P MoE Support: Introduces dedicated implementations for Mixture of Experts (MoE) models on Ascend 310P devices, including new modules for expert selection, fused MoE layers, and communication. All-Gather Communication: Enforces the use of ALLGATHER communication for MoE operations on 310P, optimizing data transfer and leveraging NPU-specific token dispatching. Simplified NPU Operations: Removes conditional type casting for npu_swiglu and enables custom rotary embedding kernels unconditionally, suggesting improved native support for 310P. New MoE Classes Registered: Registers AscendFusedMoE310 and AscendSharedFusedMoE310 to integrate 310P-specific MoE layers into the system's custom operation registry. ### Does this PR introduce _any_ user-facing change? No ### How was this patch tested? offline test and server test, with qwen3-30b-a3b,tp/ep 4 on 310p - vLLM version: v0.15.0 - vLLM main: https://github.com/vllm-project/vllm/commit/v0.15.0 --------- Signed-off-by: pu-zhe <zpuaa@outlook.com>
2026-02-06 10:30:56 +08:00
elif soc_version in {AscendDeviceType._310P}:
moe_comm_type = MoECommType.ALLGATHER
elif soc_version in {AscendDeviceType.A5}:
if num_tokens <= mc2_tokens_capacity and vllm_config.parallel_config.world_size_across_dp > 1:
moe_comm_type = MoECommType.MC2
else:
moe_comm_type = MoECommType.ALLTOALL
else:
raise ValueError(f"Unsupported soc_version: {soc_version}")
return moe_comm_type
class _ExtraForwardContextProxy:
"""Unified forward-context access for v1/v2 model runners."""
extra_attrs = (
"capturing",
"moe_comm_type",
"moe_comm_method",
"mmrs_fusion",
"num_tokens",
"flash_comm_v1_enabled",
"flashcomm_v2_enabled",
"pad_size",
"padded_length",
"num_tokens_across_dp",
"mc2_mask",
"is_draft_model",
"prefetch_mlp_gate_up_proj",
"prefetch_mlp_down_proj",
"model_instance",
"layer_idx",
"max_tokens_across_dp",
"max_tokens_across_pcp",
"num_accept_tokens",
"in_profile_run",
"padded_num_tokens",
)
def check_extra_attr(self, name: str):
if name not in self.extra_attrs:
raise AttributeError(
f"{name} is not extra forward context attribute, "
"please get/set it from vllm's _forward_context directly."
)
@staticmethod
def _ctx():
return get_forward_context()
def __getattr__(self, name: str) -> Any:
self.check_extra_attr(name)
ctx = self._ctx()
if envs_vllm.VLLM_USE_V2_MODEL_RUNNER:
return ctx.additional_kwargs[name]
return getattr(ctx, name)
def __setattr__(self, name: str, value: Any) -> None:
self.check_extra_attr(name)
ctx = self._ctx()
if envs_vllm.VLLM_USE_V2_MODEL_RUNNER:
ctx.additional_kwargs[name] = value
else:
setattr(ctx, name, value)
# usage: from vllm_ascend.ascend_forward_context import _EXTRA_CTX
_EXTRA_CTX = _ExtraForwardContextProxy()