1092626063 f63c1341d9 [Feature] GLM4.6 support mtp with fullgraph (#5460) ...

### What this PR does / why we need it?
GLM4.6 support mtp with fullgraph to improve performance

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?
`
export HCCL_BUFFSIZE=1024
export OMP_PROC_BIND=false
export OMP_NUM_THREADS=10
export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
export HCCL_OP_EXPANSION_MODE=AIV

vllm serve /weight/glm4.6_w8a8_with_float_mtp \
  --data-parallel-size 1 \
  --tensor-parallel-size 16 \
  --seed 1024 \
  --served-model-name glm \
  --max-model-len 35000 \
  --max-num-batched-tokens 16384 \
  --max-num-seqs 16 \
  --trust-remote-code \
  --gpu-memory-utilization 0.9 \
--speculative-config '{"num_speculative_tokens": 1,
"model":"/weight/glm4.6_w8a8_with_float_mtp", "method":"mtp"}' \
--compilation-config '{"cudagraph_capture_sizes": [1,2,4,8,16,32],
"cudagraph_mode": "FULL_DECODE_ONLY"}' \
  --async-scheduling \
`

test case：
`
vllm bench serve \
  --backend vllm \
  --dataset-name prefix_repetition \
  --prefix-repetition-prefix-len 22400 \
  --prefix-repetition-suffix-len 9600 \
  --prefix-repetition-output-len 1024 \
  --num-prompts 1 \
  --prefix-repetition-num-prefixes 1 \
  --ignore-eos \
  --model glm \
  --tokenizer /weight/glm4.6_w8a8_with_float_mtp \
  --seed 1000 \
  --host 0.0.0.0 \
  --port 8000 \
  --endpoint /v1/completions \
  --max-concurrency 1 \
  --request-rate 1

`
- vLLM version: v0.13.0
- vLLM main:
5326c89803

Signed-off-by: 1092626063 <1092626063@qq.com>

2026-01-09 16:07:42 +08:00

..

_cann_ops_custom

[Kernel] add custom op GmmSwigluQuantWeightNzTensorList (#3804)

2025-11-28 18:06:39 +08:00

attention

[Refactor] Replace the implementations of o_proj, q_b_proj, and kv_b_proj with custom_op for sharded CP (#5698)

2026-01-09 15:58:40 +08:00

compilation

[BugFix][Fusion] Fix graph fusion failure problem (#5676)

2026-01-07 18:42:55 +08:00

core

[CI] fix lint (#5216)

2025-12-20 17:03:25 +08:00

device_allocator

[Refactor] Cleanup platform (#5566)

2026-01-07 09:25:55 +08:00

distributed

[bugfix] adapt to new implemented get_kv_cache_spec in cpuoffload connector (#4311)

2026-01-08 09:15:09 +08:00

eplb

[Bugfix] Revert pr4214 multi-stream collect expert hotpot (#5529)

2026-01-07 11:26:47 +08:00

kv_offload

[BugFix] Fix npu-cpu offloading interface change bug. (#5290)

2025-12-27 10:21:20 +08:00

lora

[BufFix]Fix the error when using Ascend custom operators with rank=128 (#5394)

2026-01-09 15:57:43 +08:00

model_loader

[BugFix] NetLoader: No backend type associated with device type npu (#5700)

2026-01-09 15:54:54 +08:00

ops

[Bugfix] Fix matmul allreduce precision issue by using original weight (#4939)

2026-01-09 16:05:32 +08:00

patch

[BugFix][Fusion] Fix graph fusion failure problem (#5676)

2026-01-07 18:42:55 +08:00

quantization

[Feature] GLM4.6 support mtp with fullgraph (#5460)

2026-01-09 16:07:42 +08:00

sample

[Feature] add the magicmtp speculative decoding acceleration algorithm (#5542)

2026-01-08 09:15:55 +08:00

spec_decode

[feature]dcp&pcp support mlapo (#5672)

2026-01-08 23:49:23 +08:00

worker

[feature]dcp&pcp support mlapo (#5672)

2026-01-08 23:49:23 +08:00

xlite

[BugFix] Xlite: Bypass the padding of the graph mode in non-MTP cases to obtain the correct decode num. (#5711)

2026-01-09 15:55:30 +08:00

__init__.py

clean up model module (#4611)

2025-12-02 17:35:47 +08:00

ascend_config.py

[refactor] Refactor the interface for shard weight and remove the flashcomm2 o_shared interface. (#5181)

2026-01-08 09:05:02 +08:00

ascend_forward_context.py

[Feature]EPLB:Adapt DispatchGmmCombineDecode operator to eplb tensor list and expert token numbers (#5552)

2026-01-07 11:23:42 +08:00

batch_invariant.py

[Feature] implement basic framework for batch invariant (#5517)

2026-01-07 09:11:26 +08:00

cpu_binding.py

[main] support cpu binding (#3546)

2025-10-21 09:17:03 +08:00

envs.py

[refactor] Refactor the interface for shard weight and remove the flashcomm2 o_shared interface. (#5181)

2026-01-08 09:05:02 +08:00

flash_common3_context.py

[Perf]enable prefill flashcommon3 (#4065)

2025-12-14 09:34:13 +08:00

meta_registration.py

Fix the bugs about operator registration by PyTorch Dispatcher (#2786)

2025-09-13 11:58:52 +08:00

platform.py

Optimize the print info format when deprecated code is used in vllm-ascend (#5696)

2026-01-08 09:26:49 +08:00

profiling_config.py

Drop ascend scheduler (#4623)

2025-12-05 09:03:45 +08:00

utils.py

[Refactor] Replace the implementations of o_proj, q_b_proj, and kv_b_proj with custom_op for sharded CP (#5698)

2026-01-09 15:58:40 +08:00