fb0d6dd175
### What this PR does / why we need it?
Fixed the error of speculative decoding in FULL mode when `num_spec + 1`
not in `cudagraph_capture_sizes`.
Now, we can run speculative decoding in FULL mode, but with drafter as
eager.
It depends on https://github.com/vllm-project/vllm-ascend/pull/7144 .
### Does this PR introduce _any_ user-facing change?
N/A
### How was this patch tested?
Test code is shown as below:
```python
prompts = [
"1.Who are you?",
"2. Who are you?",
]
sampling_params = SamplingParams(temperature=0.0, top_p=0.95, top_k=40, max_tokens=200)
llm = LLM(
model="/home/some-model/Meta-Llama-3.1-8B-Instruct",
tensor_parallel_size=1,
max_num_seqs=32,
# enforce_eager=True,
disable_log_stats=False,
distributed_executor_backend="mp",
gpu_memory_utilization=0.7,
async_scheduling=True,
speculative_config={
"enforce_eager": True,
"model": "/home/some-model/EAGLE3-LLaMA3.1-Instruct-8B",
"disable_padded_drafter_batch": False,
"method": "eagle3",
"num_speculative_tokens": 2,
},
compilation_config={
"cudagraph_mode": "FULL",
"cudagraph_num_of_warmups": 1,
},
max_model_len=4096,
enable_prefix_caching=False,
)
outputs = llm.generate(prompts, sampling_params)
```
The result before:
```text
File "/vllm-workspace/vllm/vllm/v1/cudagraph_dispatcher.py", line 140, in _create_padded_batch_descriptor
assert num_tokens_padded % uniform_decode_query_len == 0
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
AssertionError
```
The result after:
```text
--------------------------------------------------
total_num_output_tokens: 400
num_drafts: 249
num_draft_tokens: 498
num_accepted_tokens: 149
mean acceptance length: 1.60
--------------------------------------------------
acceptance at token 0: 0.43
acceptance at token 1: 0.17
```
- vLLM version: v0.16.0
- vLLM main:
4034c3d32e
Signed-off-by: drslark <slarksblood@qq.com>
455 lines
21 KiB
Python
455 lines
21 KiB
Python
#
|
||
# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
|
||
# This file is a part of the vllm-ascend project.
|
||
#
|
||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||
# you may not use this file except in compliance with the License.
|
||
# You may obtain a copy of the License at
|
||
#
|
||
# http://www.apache.org/licenses/LICENSE-2.0
|
||
#
|
||
# Unless required by applicable law or agreed to in writing, software
|
||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||
# See the License for the specific language governing permissions and
|
||
# limitations under the License.
|
||
|
||
# ----------------------------------------------------------------------------------
|
||
# This module manage the patch for vllm. There are two folders in this module:
|
||
# - platform: contains the patches applied before worker starts. It's called by
|
||
# `vllm_ascend.utils.adapt_patch(is_global_patch=True)` in
|
||
# `vllm_ascend.platform.NPUPlatform.pre_register_and_update()` function.
|
||
# - worker: contains the patches applied when worker starts. It's called by
|
||
# `vllm_ascend.utils.adapt_patch(is_global_patch=False)` in
|
||
# each worker's `__init__` function.
|
||
#
|
||
# Once a new patch is added in vllm-ascend, please add the patch description into this file as well.
|
||
# ----------------------------------------------------------------------------------
|
||
|
||
# What's Patched and how it works:
|
||
# --------------------------------
|
||
# * Platform Patch:
|
||
# =================
|
||
# ** 1. File: platform/patch_distributed.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `torch.distributed.all_reduce`, `torch.distributed.broadcast`
|
||
# Why:
|
||
# tensor alignment for 310p
|
||
# How：
|
||
# rewrite all_reduce and broadcast in torch.distributed
|
||
# Related PR (if no, explain why):
|
||
# No, not ready yet.
|
||
# Future Plan:
|
||
# Find a better way to support tensor alignment for 310p without this patch.
|
||
#
|
||
# ** 2. File: platform/patch_mamba_config.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.config.HybridAttentionMambaModelConfig.verify_and_update_config`
|
||
# Why:
|
||
# block size is set to 16 in vLLM which is not supported by Ascend.
|
||
# How：
|
||
# Set block size to 128 on npu.
|
||
# Related PR (if no, explain why):
|
||
# we'll fix this in vLLM soon.
|
||
# Future Plan:
|
||
# Remove this patch when vLLM merges the PR.
|
||
#
|
||
# ** 3. File: platform/patch_multiproc_executor.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.executor.multiproc_executor.MultiprocExecutor`
|
||
# Why:
|
||
# vLLM create child process with daemon=True, which doesn't work with EPLB case, since EPLB will create
|
||
# a new process which is not allowed by daemon=True.
|
||
# How：
|
||
# Set daemon=False in MultiprocExecutor.
|
||
# Related PR (if no, explain why):
|
||
# Find a way to support daemon=False in vLLM
|
||
# Future Plan:
|
||
# Remove this patch when vLLM fix the issue.
|
||
#
|
||
# ** 4. File: platform/patch_sched_yield.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.distributed.utils.USE_SCHED_YIELD`
|
||
# Why:
|
||
# os.sched_yield() doesn't work on Arm systems.
|
||
# How：
|
||
# avoid using os.sched_yield() on Arm systems.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/30228
|
||
# Future Plan:
|
||
# Remove this patch when vLLM merge the PR.
|
||
#
|
||
# ** 5. File: platform/patch_balance_schedule.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.engine.core.EngineCoreProc.run_engine_core`
|
||
# `vllm.v1.core.sched.scheduler.Scheduler`
|
||
# Why:
|
||
# vLLM v1 scheduling currently enables chunkedprefill by default, which processes prefill and decode
|
||
# requests simultaneously in a single scheduling session. This can impact the overall system throughput
|
||
# and performance in some scenarios.
|
||
# How：
|
||
# Set environmental variables VLLM_ASCEND_BALANCE_SCHEDULING=1 in startup script.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/29721
|
||
# Future Plan:
|
||
# Remove this patch when vLLM merge the PR.
|
||
#
|
||
# ** 6. File: platform/patch_fusion_matcher_compat_ops.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `torch.ops._C.rms_norm`, `torch.ops._C.fused_add_rms_norm`,
|
||
# Why:
|
||
# upstream vLLM initializes fusion matcher global operators at import time.
|
||
# On Ascend environment these symbols may be absent and cause import failure.
|
||
# How：
|
||
# inject placeholders only when the symbols are missing so import can continue.
|
||
# Related PR (if no, explain why):
|
||
# temporary compatibility patch before upstream adjustment is merged.
|
||
# Future Plan:
|
||
# remove this patch once upstream no longer requires these global symbols or
|
||
# provides a backend-safe initialization path.
|
||
#
|
||
# ** 7. File: platform/patch_minimax_m2_config.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.config.model.ModelConfig._verify_quantization`
|
||
# Why:
|
||
# MiniMax-M2 fp8 checkpoints on NPU may fail upstream quantization validation.
|
||
# vllm-ascend needs to disable fp8 quantization and load bf16 dequantized
|
||
# weights in worker-side patches instead.
|
||
# How：
|
||
# Monkey-patch `_verify_quantization` and intercept platform quantization
|
||
# verification to force `cfg.quantization=None` for MiniMax-M2 fp8 on NPU.
|
||
# Related PR (if no, explain why):
|
||
# No, upstream behavior differs across versions and needs discussion.
|
||
# Future Plan:
|
||
# Remove this patch once upstream supports MiniMax-M2 fp8 on NPU or provides
|
||
# a backend-safe validation / override mechanism.
|
||
#
|
||
# 2. `vllm.config.model.ModelConfig._verify_cuda_graph`
|
||
# Why:
|
||
# For MiniMax-M2 on NPU with ACL graph capture enabled, HCCL op expansion
|
||
# mode affects graph shape coverage. Users may forget to set it.
|
||
# How：
|
||
# If user doesn't set it, set `HCCL_OP_EXPANSION_MODE=AIV` for this model
|
||
# and log a warning when a different value is detected.
|
||
# Related PR (if no, explain why):
|
||
# No, this is an environment-specific tuning knob.
|
||
# Future Plan:
|
||
# Remove this patch if upstream provides an official NPU graph-capture
|
||
# guidance / auto-configuration path for HCCL.
|
||
#
|
||
# * Worker Patch:
|
||
# ===============
|
||
#
|
||
# ** 1. File: worker/patch_distributed.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.distributed.parallel_state.GroupCoordinator`
|
||
# Why:
|
||
# vllm doesn't support all_to_all for GroupCoordinator.
|
||
# all_reduce in vLLM not is a customop, which will make MatmulAllReduceAddRMSNorm fusion failure.
|
||
# How：
|
||
# Add all_to_all implementation for GroupCoordinator.
|
||
# make all_reduce as a customop.
|
||
# Related PR (if no, explain why):
|
||
# No, we should use vlLM all2all manager to support all_to_all for npu.
|
||
# Future Plan:
|
||
# Remove this patch when the refactor of all2all manager is done.
|
||
# Remove this patch when vLLM support all_reduce as customop.
|
||
#
|
||
# ** 2. File: worker/patch_multimodal_merge.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.utils._merge_multimodal_embeddings`
|
||
# Why:
|
||
# '_merge_multimodal_embeddings' func of vllm is incompatible with Ascend.
|
||
# How：
|
||
# Replace with CPU operation that can be executed asynchronously.
|
||
# Related PR (if no, explain why):
|
||
# This is a bug by Ascend only. It can' be fixed in vLLM.
|
||
# Future Plan:
|
||
# Identify this pattern in torch-npu and remove this patch.
|
||
#
|
||
# ** 3. File: worker/patch_bert.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.bert._encode_token_type_ids`
|
||
# `vllm.model_executor.models.bert._decode_token_type_ids`
|
||
# Why:
|
||
# shift operation in `_encode_token_type_ids` and `_decode_token_type_ids` cannot run in ascend aclgraph mode
|
||
# How：
|
||
# Replace shift operation with multiplication and division.
|
||
# Related PR (if no, explain why):
|
||
# No, this need CANN add an aclnn shift operation
|
||
# Future Plan:
|
||
# Revert this when CANN support shift aclnn operation
|
||
#
|
||
# ** 4. File: worker/patch_triton.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.layers.mamba.ops`, `vllm.model_executor.layers.fla.ops`,
|
||
# `vllm.v1.worker.gpu.sample.gumbel.gumbel_sample`
|
||
# Why:
|
||
# triton ops in vLLM perform not good on NPU. And there is no dispatch mechanism for triton ops.
|
||
# How：
|
||
# override triton ops in vLLM with ascend implementation
|
||
# Related PR (if no, explain why):
|
||
# Let vLLM support triton ops dispatch.
|
||
# Future Plan:
|
||
# Remove this patch when vLLM support the dispatch function.
|
||
#
|
||
# ** 5. File: worker/patch_qwen3_next_mtp.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.worker.utils.bind_kv_cache`
|
||
# Why:
|
||
# 'bind_kv_cache' func will raise an exception when current_platform is npu.
|
||
# How：
|
||
# Replace with a new bind_kv_cache.
|
||
# Skip the raise.
|
||
# Related PR (if no, explain why):
|
||
# It need discuss.
|
||
# Future Plan:
|
||
# Remove this patch after discussing with vllm community and adapting bind_kv_cache to npu.
|
||
#
|
||
# ** 6. File: worker/patch_rejection_sampler.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.sample.rejection_sampler`
|
||
# Why:
|
||
# - some functions from `rejection_sampler` are not supported or slow on npu.
|
||
# How：
|
||
# - add npu_top_k_top_p to 'apply_sampling_constraints' func
|
||
# - add custom triton kernel to `expand_batch_to_tokens` and `rejection_sample`
|
||
# Related PR (if no, explain why):
|
||
# Let vLLM support triton ops dispatch.
|
||
# Future Plan:
|
||
# 1. make these functions as class func of RejectionSampler, create AscendRejectionSampler
|
||
# to override them, then delete the patch file `worker/patch_rejection_sampler.py`.
|
||
# 2. make these functions as costom op, then remove AscendRejectionSampler
|
||
#
|
||
## ** 7. File: worker/patch_module.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.attention.backends.gdn_attn.torch.argsort`
|
||
# Why:
|
||
# 1. 'torch.argsort' func of npu does not support bool.
|
||
# 2. Without `stable=True`, the output will have a lot of redundant tokens.
|
||
# How：
|
||
# Replace with a new torch.argsort that will cast the input to torch.int32
|
||
# and do stable sort.
|
||
# Related PR (if no, explain why):
|
||
# 1. It depends on torch_npu.
|
||
# 2. https://github.com/vllm-project/vllm/pull/30632
|
||
# Future Plan:
|
||
# Remove this patch when bool is supported in 'torch.argsort' func of npu.
|
||
# Make 'torch.argsort' in `vllm.v1.attention.backends.gdn_attn` be stable.
|
||
#
|
||
# ** 8. File: worker/patch_qwen3_next.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.qwen3_next.Qwen3NextGatedDeltaNet.forward`
|
||
# Why:
|
||
# The Qwen3Next GatedDeltaNet forward cannot directly add custom operators.
|
||
# How：
|
||
# Add a branch in Qwen3NextGatedDeltaNet.forward to adapt to fused_qkvzba_split_reshape_cat.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/30863
|
||
# Future Plan:
|
||
# Remove this patch when vLLM support these operators.
|
||
#
|
||
# 2. `vllm.model_executor.models.qwen3_next.Qwen3NextGatedDeltaNet._forward_core`
|
||
# Why:
|
||
# triton ops fused_recurrent_gated_delta_rule and fused_gdn_gating in vLLM perform not good on NPU.
|
||
# How：
|
||
# add a new fused triton ops in vLLM with ascend implementation.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/30860
|
||
# Future Plan:
|
||
# Remove this patch when vLLM support these operators.
|
||
#
|
||
# 3. `vllm.model_executor.models.qwen3_next.Qwen3NextGatedDeltaNet._forward_core`
|
||
# Why:
|
||
# The Qwen3Next GatedDeltaNet _forward_core cannot directly add custom operators.
|
||
# How：
|
||
# Add a branch in Qwen3NextGatedDeltaNet._forward_core to adapt to fused_gdn_gating_patch.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/31002
|
||
# Future Plan:
|
||
# Remove this patch when vLLM support these operators.
|
||
#
|
||
# ** 9. File: worker/patch_huanyuan_vl.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.transformers_utils.processors.hunyuan_vl.HunYuanVLProcessor.__call__`
|
||
# Why:
|
||
# The `add_special_tokens` parameter is not supported by default in the processor.
|
||
# How：
|
||
# Remove the `add_special_tokens` parameter from kwargs before calling the original method.
|
||
# Future Plan:
|
||
# Remove this patch when vLLM aligns with the latest processor implementation.
|
||
#
|
||
# ** 10. File: worker/patch_v2_eagle.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.worker.gpu.spec_decode.eagle.EagleSpeculator.propose`
|
||
# Why:
|
||
# `propose` method use torch.gather, but the gather operator will
|
||
# pollute the arguments passed to it. the bug is reported to huawei
|
||
# CANN team, but not fixed yet.
|
||
# How：
|
||
# clone the out attribute ahead of gather to avoid the bug.
|
||
# Future Plan:
|
||
# Remove this patch when cann fix the gather bug.
|
||
#
|
||
# ** 11. File: worker/patch_unquantized_gemm.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.layers.utils.default_unquantized_gemm`
|
||
# Why:
|
||
# unquantized_gemm in vLLM not is a customop, which will make MatmulAllReduceAddRMSNorm fusion failure.
|
||
# How：
|
||
# make unquantized_gemm as a customop.
|
||
# Future Plan:
|
||
# Remove this patch when vLLM support the operator as customop.
|
||
#
|
||
# ** 12. File: worker/patch_npugraph_ex_triton.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `torchair.core._concrete_graph.ValuePack`,
|
||
# `torchair.npu_fx_compiler._unpack_meta`,
|
||
# `torchair.npu_fx_compiler._NpuGraphConverter._unpack_npu`
|
||
# Why:
|
||
# In the Triton scenario, npugraph_ex backend needs to process the value pack of the input parameters.
|
||
# How：
|
||
# Supplement the relevant processing logic through patches.
|
||
# Related PR (if no, explain why):
|
||
# https://gitcode.com/Ascend/torchair/pull/2575
|
||
# Future Plan:
|
||
# Remove this patch when the PTA version used by vllm-ascend has been upgraded.
|
||
#
|
||
# ** 13. File: worker/patch_v2_uva.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.worker.gpu.states.UvaBuffer`
|
||
# Why:
|
||
# ASCEND NPUs do not support UVA yet, so we need to wrap it in vLLM.
|
||
# How：
|
||
# make UvaBuffer a dummy class, mimic the interface of vllm UvaBuffer.
|
||
# Future Plan:
|
||
# Remove this patch when NPU support UVA.
|
||
#
|
||
# ** 14. File: worker/patch_kimi_k25.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.kimi_k25_vit.Learnable2DInterpPosEmbDivided_fixed.forward`
|
||
# Why:
|
||
# The forward method uses interpolate with ops not supported on NPU.
|
||
# How：
|
||
# Replace with a new forward that uses CPU for interpolate when shape mismatch,
|
||
# and use get_rope_shape to handle the rope shape interpolation.
|
||
# Future Plan:
|
||
# Remove this patch when vLLM aligns with the latest main.
|
||
#
|
||
# ** 15. File: worker/patch_routed_experts_capturer.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.layers.fused_moe.routed_experts_capturer.RoutedExpertsCapturer.init_buffer`
|
||
# Why:
|
||
# The `_device_buffer` initialization in vLLM uses `device="cuda"` hardcoded,
|
||
# which doesn't work on NPU.
|
||
# How：
|
||
# Replace `device="cuda"` with `device=current_platform.device_name` to support NPU.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/34336
|
||
# Future Plan:
|
||
# Remove this patch when vLLM merges the PR.
|
||
# ** 16. File: worker/patch_draft_quarot.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.llama_eagle3.Eagle3LlamaForCausalLM.load_weights`
|
||
# Why:
|
||
# vllm-ascend reused the loading logic of drafter model from vllm,
|
||
# but vllm doesn't need to apply to Ascend quantization.
|
||
# How：
|
||
# Dynamically replace the `load_weights` function at runtime,
|
||
# and fix `target_config` into the new implementation with a closure.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/36225
|
||
# Future Plan:
|
||
# Remove this patch when vLLM merges the PR.
|
||
#
|
||
# ** 17. File: worker/patch_minimax_m2.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.minimax_m2.MiniMaxM2MoE.forward`
|
||
# Why:
|
||
# In TP mode, MiniMax-M2 MoE needs a backend-aware reduction path to avoid
|
||
# unnecessary communication / maintain correctness on NPU.
|
||
# How：
|
||
# Replace the forward to call `experts.maybe_all_reduce_tensor_model_parallel`
|
||
# when `tp_size > 1`.
|
||
# Related PR (if no, explain why):
|
||
# No, model-specific behavior.
|
||
# Future Plan:
|
||
# Move this behavior upstream once a generic MoE reduce hook exists.
|
||
#
|
||
# 2. `vllm.model_executor.models.minimax_m2.MiniMaxM2Attention.__init__`
|
||
# Why:
|
||
# When total kv heads < TP world size, kv head replication happens and k_norm
|
||
# weights should be sharded to match the replication layout.
|
||
# How：
|
||
# Add `num_kv_head_replicas` and create sharded `k_norm` via
|
||
# `MiniMaxText01RMSNormTP(..., weight_shard_world_size=total_num_kv_heads, ...)`.
|
||
# Related PR (if no, explain why):
|
||
# No, depends on Ascend kernel behavior and TP layout.
|
||
# Future Plan:
|
||
# Remove this patch if upstream implements kv-head-aware norm sharding.
|
||
#
|
||
# 3. `vllm.model_executor.models.minimax_m2.MiniMaxM2Model.load_weights`
|
||
# Why:
|
||
# MiniMax-M2 fp8 checkpoints may store fp8 weights with per-block inverse
|
||
# scales. On NPU we load bf16 weights by dequantizing at load time.
|
||
# How：
|
||
# Inject fp8 dequant helpers and wrap `load_weights` to convert fp8 weight +
|
||
# `weight_scale_inv` pairs into bf16 blocks before delegating to upstream.
|
||
# Related PR (if no, explain why):
|
||
# No, fp8 load format and backend constraints are model/backend specific.
|
||
# Future Plan:
|
||
# Remove this patch when upstream supports MiniMax-M2 fp8 loading on NPU.
|
||
#
|
||
# ** 18. File: worker/patch_minimax_m2_linear_attn.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.layers.mamba.linear_attn.MiniMaxText01RMSNormTP.__init__`
|
||
# `vllm.model_executor.layers.mamba.linear_attn.MiniMaxText01RMSNormTP.weight_loader`
|
||
# Why:
|
||
# MiniMax-M2 linear attention RMSNorm needs weight sharding that can follow
|
||
# TP layout (and sometimes kv-head replication) on NPU.
|
||
# How：
|
||
# Override `__init__` to parameterize weight shard world/rank and install a
|
||
# sharded `weight_loader` implementation.
|
||
# Related PR (if no, explain why):
|
||
# No, upstream API surface differs across versions.
|
||
# Future Plan:
|
||
# Remove this patch when upstream exposes stable sharding hooks for this layer.
|
||
#
|
||
# 2. `vllm.model_executor.layers.mamba.linear_attn.MiniMaxText01RMSNormTP.forward_qk`
|
||
# (or older `_normalize_qk`)
|
||
# Why:
|
||
# q/k norm for linear attention is performance-sensitive. On NPU, a fused
|
||
# rms_norm kernel is faster and TP needs a global rstd correction.
|
||
# How：
|
||
# Replace q/k normalization with NPU rms_norm fast path and TP-global rstd
|
||
# correction; fall back to upstream implementation on non-NPU.
|
||
# Related PR (if no, explain why):
|
||
# No, backend-specific optimization.
|
||
# Future Plan:
|
||
# Remove this patch when upstream adds a backend dispatch path for q/k norm.
|
||
#
|
||
# ** 19. File: worker/patch_qwen3_5.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.model_executor.models.qwen3_5.Qwen3_5GatedDeltaNet._forward_core`
|
||
# Why:
|
||
# The class Qwen3_5GatedDeltaNet reuse the `_forward_core` method of Qwen3NextGatedDeltaNet,
|
||
# but the ascendC ops of Qwen3NextGatedDeltaNet do not support ssm_state with float32 format.
|
||
# How：
|
||
# patch Qwen3_5GatedDeltaNet._forward_core to use triton ops like `fused_recurrent_gated_delta_rule`.
|
||
# Future Plan:
|
||
# Remove this patch when all ops in _forward_core support both Qwen3_5 and Qwen3Next.
|
||
#
|
||
# ** 20. File: worker/patch_cudagraph.py**
|
||
# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
|
||
# 1. `vllm.v1.cudagraph_dispatcher.CudagraphDispatcher._create_padded_batch_descriptor`
|
||
# Why:
|
||
# vllm's FULL mode will cause error, we use a patch to avoid it.
|
||
# After that, FULL can be enable now.
|
||
# How：
|
||
# Dynamically replace the `_create_padded_batch_descriptor` function at runtime,
|
||
# and change the condition of if.
|
||
# Related PR (if no, explain why):
|
||
# https://github.com/vllm-project/vllm/pull/34880
|
||
# Future Plan:
|
||
# Remove this patch when vLLM merges the PR.
|