[Feature][Model] Add Kimi K3 support for vLLM 0.27

.github/workflows/scripts/runner_label.json

@@ -35,6 +35,12 @@
         "image_tag": "9.1.0-a3-ubuntu22.04-py3.12",
         "csrc_cache_target": "a3-arm64-ubuntu"
     },
+    "linux-aarch64-a3-16": {
+        "chip": "a3",
+        "npu_num": 16,
+        "image_tag": "9.1.0-a3-ubuntu22.04-py3.12",
+        "csrc_cache_target": "a3-arm64-ubuntu"
+    },
     "linux-aarch64-a3-2-": {
         "chip": "a3",
         "npu_num": 2,

.github/workflows/scripts/test_config.yaml

@@ -308,6 +308,7 @@
     - tests/ut/models/test_deepseek_v4_compressor.py
     - tests/ut/models/test_deepseek_v4_indexer.py
     - tests/ut/models/test_deepseek_v4_moe.py
+    - tests/ut/models/test_kimi_k3_adapter.py
     - tests/e2e/pull_request/four_card/test_deepseek_v4.py
 
 - name: models_minimax_m3
@@ -321,6 +322,30 @@
     - tests/e2e/pull_request/one_card/test_minimax_m3_sparse_attn.py
     - tests/e2e/pull_request/eight_card/test_minimax_m3.py
 
+- name: models_kimi_k3
+  optional: true
+  source_file_dependencies:
+    - vllm_ascend/models/kimi_k3.py
+    - vllm_ascend/models/kimi_k3_dspark.py
+    - vllm_ascend/models/kimi_k3_mtp.py
+    - vllm_ascend/models/qwen3_dspark.py
+    - vllm_ascend/worker/model_runner_v1.py
+    - vllm_ascend/worker/worker.py
+    - vllm_ascend/attention/mla_v1.py
+    - vllm_ascend/core/kv_cache_interface.py
+    - vllm_ascend/ops/kimi_kda.py
+    - vllm_ascend/ops/fused_moe
+    - vllm_ascend/ops/triton/kimi_k3
+    - vllm_ascend/spec_decode/dspark_proposer.py
+    - vllm_ascend/spec_decode/llm_base_proposer.py
+    - vllm_ascend/quantization/modelslim_config.py
+    - vllm_ascend/patch/platform/patch_kv_cache_utils.py
+    - vllm_ascend/patch/platform/patch_kv_cache_coordinator.py
+    - vllm_ascend/patch/platform/patch_speculative_config.py
+    - vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py
+  tests:
+    - tests/e2e/pull_request/sixteen_card/test_kimi_k3.py
+
 # Ops
 - name: ops_basic
   optional: false
@@ -391,6 +416,7 @@
   source_file_dependencies:
     - vllm_ascend/ops/gdn.py
     - vllm_ascend/ops/gdn_attn_builder.py
+    - vllm_ascend/ops/kimi_kda.py
     - vllm_ascend/ops/triton/gdn_chunk_meta.py
   tests:
     - tests/ut/ops
@@ -892,6 +918,7 @@ estimated_times:
   tests/e2e/pull_request/four_card/test_qwen3_32b_bf16_ll_performance.py: 1200
   tests/e2e/pull_request/two_card/test_gemma4.py: 420
   tests/e2e/pull_request/eight_card/test_minimax_m3.py: 630
+  tests/e2e/pull_request/sixteen_card/test_kimi_k3.py: 360
   tests/ut/attention/a2/test_attention_cp.py: 30
   tests/ut/attention/a2/test_attention_cp_precision.py: 30
   tests/ut/attention/a2/test_attention_v1.py: 30
@@ -913,6 +940,7 @@ estimated_times:
   tests/ut/ops/a2/test_gdn_layerwise_kv.py: 70
   tests/ut/ops/a2/test_token_dispatcher.py: 30
   tests/ut/ops/a3_2/test_activation.py: 50
+  tests/ut/ops/a3_2/test_kimi_kda_fused_norm_gate.py: 60
   tests/ut/ops/a3_2/test_select_experts.py: 20
   tests/ut/quantization/methods/a2/test_w4a16.py: 30
   tests/ut/quantization/methods/a2/test_w4a4_flatquant.py: 40
@@ -971,6 +999,8 @@ runner_mapping:
     310p: 310p-4
   tests/e2e/pull_request/eight_card:
     default: a3-8
+  tests/e2e/pull_request/sixteen_card:
+    default: a3-16
   tests/ut/.+/a3_2:
     default: a3-2
   tests/ut/.+/a2:
@@ -1004,6 +1034,9 @@ partition:
   a3-8:
     runner_label: linux-aarch64-a3-8-
     count: 1
+  a3-16:
+    runner_label: linux-aarch64-a3-16
+    count: 1
   cpu-0:
     runner_label: linux-amd64-cpu-8-hk
     count: 1

.gitleaks.toml

@@ -56,7 +56,8 @@ paths = ["^vllm-empty/"]
 # Allow tilingKey generic-api-key false positives in the specified file
 [[allowlists]]
 paths = [
-    "^csrc/notify_dispatch/op_host/notify_dispatch_tiling.cpp"
+    "^csrc/notify_dispatch/op_host/notify_dispatch_tiling.cpp",
+    "^csrc/attention/mla_prolog_v3/op_kernel/mla_prolog_template_tiling_key.h"
 ]
 rules = ["generic-api-key"]
 

.pre-commit-config.yaml

@@ -22,7 +22,7 @@ repos:
       args: [
         --toml, pyproject.toml,
         '--skip', 'tests/prompts/**,./benchmarks/sonnet.txt,*tests/lora/data/**,build/**,./vllm_ascend.egg-info/**,typos.toml',
-        '-L', 'CANN,cann,NNAL,nnal,ASCEND,ascend,EnQue,CopyIn,ArchType,AND,ND,tbe,copyin,alog,outter,mata,PARD'
+        '-L', 'CANN,cann,NNAL,nnal,ASCEND,ascend,EnQue,CopyIn,ArchType,AND,ND,tbe,copyin,alog,outter,mata,PARD,uSeed,LoadIn'
       ]
       additional_dependencies:
         - tomli

csrc/attention/chunk_kda_fwd/CMakeLists.txt

@@ -0,0 +1,16 @@
+# -----------------------------------------------------------------------------------------------------------
+# Copyright (c) 2026 Tianjin University, Ltd.
+# This program is free software, you can redistribute it and/or modify it under the terms and conditions of
+# Please refer to the License for details. You may not use this file except in compliance with the License.
+# THIS SOFTWARE IS PROVIDED ON AN "AS IS" BASIS, WITHOUT WARRANTIES OF ANY KIND, EITHER EXPRESS OR IMPLIED,
+# INCLUDING BUT NOT LIMITED TO NON-INFRINGEMENT, MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.
+# -----------------------------------------------------------------------------------------------------------
+file(GLOB CURRENT_DIRS RELATIVE ${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_SOURCE_DIR}/*)
+if(NOT ENABLE_TEST AND NOT BENCHMARK)
+    list(REMOVE_ITEM CURRENT_DIRS tests)
+endif()
+foreach(SUB_DIR ${CURRENT_DIRS})
+    if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/${SUB_DIR}/CMakeLists.txt")
+        add_subdirectory(${SUB_DIR})
+    endif()
+endforeach()

csrc/attention/chunk_kda_fwd/README.md

@@ -0,0 +1,114 @@
+# ChunkKdaFwd
+
+## 功能
+
+`ChunkKdaFwd` 对齐不涉及 CP 切分的 FLA `chunk_kda_fwd` 顶层语义。公共接口接收 raw gate 或已激活的
+自然对数 gate；Gate、Prepare、PostWu、FwdH 和 Finalize 均在一个物理 `ChunkKdaFwd` L0 内完成，
+L2 不再拼接或依次发射多个阶段 L0。
+
+Shape 符号与布局约定见 [KDA 模型符号表](../README.md#model-shape-symbols)。
+
+## Gate 公式
+
+令 `x = g + dt_bias`。逐 token、逐 K 维的自然对数衰减为：
+
+```text
+use_gate_in_kernel = false:
+    gate = g
+
+use_gate_in_kernel = true, safe_gate = false:
+    gate = -exp(A_log) * softplus(x)
+
+use_gate_in_kernel = true, safe_gate = true:
+    gate = lower_bound * sigmoid(exp(A_log) * x)
+```
+
+随后在每个 chunk 内计算：
+
+```text
+gk_i = cumsum(gate)_i / ln(2)
+```
+
+因此后续 `exp2(gk)` 与自然指数 gate 严格绑定，不暴露额外 gate scale。
+
+## 输入
+
+| 名称 | 必选性 | Shape/Dtype | 说明 |
+| --- | --- | --- | --- |
+| `q/k` | 必选 | 输入 layout 对应 Shape；FP16/BF16 | Query/Key |
+| `v` | 必选 | 输入 layout 对应 Shape；与 q 同 dtype | Value |
+| `g` | 必选 | 输入 layout 对应 K 维 Shape；FP32/BF16 | raw gate 或已激活自然对数 gate |
+| `beta` | 必选 | 去掉 g 的 K 维；FP32/BF16 | Delta 系数 |
+| `A_log` | 条件必选 | `[H_v]`，FP32 | `use_gate_in_kernel=true` 时必选 |
+| `dt_bias` | 可选 | `[H_v*K]`，FP32 | gate bias |
+| `initial_state` | 可选 | `[N,H_v,K,V]` 或 `[N,H_v,V,K]`，FP32 | 由 `state_v_first` 解释 |
+| `cu_seqlens` | 可选 | `[N+1]`，INT64 | 变长序列 |
+| `chunk_indices` | 可选 | `[2*N_c]`，INT64 | canonical chunk 顺序 |
+
+`layout` 只描述上述输入。BSND/TND 由 L2 使用 `l0op::Transpose` 转为内部 BNSD/NTD。
+
+## 输出
+
+Python 返回顺序为：
+
+```text
+(attn_out, final_state, gk, Aqk, Akk, w, u, qg, kg, v_new, h, initial_state)
+```
+
+- `attn_out` 固定为 BSND/TND。
+- `final_state` 固定按序列排列，末两维服从 `state_v_first`。
+- `Aqk/Akk` 始终返回，固定为 head-major。
+- `gk/w/u/qg/kg/v_new` 是供反向使用的 head-major 中间量。
+- 公开 `h` 固定为 sequence-major；内部 `hCompute` 保持 head-major 供 Finalize 使用。
+- 第 12 个返回值是 Python 层对 `initial_state` 的原对象透传，不是 aclnn 输出。
+
+输出保留策略对齐 fla-org
+[`chunk_kda_fwd`](https://github.com/fla-org/flash-linear-attention/blob/0f0f0c97af39343855b43bbbaddcedfda5cb9d77/fla/ops/kda/chunk_fwd.py)
+提交 `0f0f0c97af39343855b43bbbaddcedfda5cb9d77`：
+
+| 条件 | 返回 |
+| --- | --- |
+| `output_final_state=true` | 返回 `final_state`，否则为 `None` |
+| `use_gate_in_kernel=false` 或 `disable_recompute=true` | 返回 `gk` |
+| 始终 | 返回 `Aqk/Akk` |
+| `disable_recompute=true` | 返回 `w/u/qg/kg/v_new` |
+| `disable_recompute=true` 或 `return_intermediate_states=true` | 返回 `h` |
+
+这是 `fla_npu.ops.ascendc.chunk_kda_fwd` 的低层 12 返回值语义；不涉及 CP。aclnn L2 不接收
+`output_final_state/disable_recompute/return_intermediate_states`，每个可选输出是否写出仅由对应
+输出指针是否为空决定。`w/u/qg/kg/v_new/h` 的 L0 阶段固定写内部 compute 张量，L2 仅在
+对应指针非空时通过 `ViewCopy` 导出；`gkOut` 非空时直接复用为 `gkCompute`，避免目标场景
+额外复制整张 FP32 gate。内部 `hCompute` 是 FwdH 到 Finalize 的必需 head-major 阶段结果；
+公开 `hOut` 非空时，L2 转为 sequence-major 后导出。`hOut` 为空时仍创建 `hCompute`，但不
+作为第 11 个 Python 返回值公开。
+
+## 属性
+
+| 名称 | 默认值 | 支持范围 |
+| --- | --- | --- |
+| `layout` | `BSND` | `BSND/BNSD/TND/NTD` |
+| `scale` | 必传 | 通常为 `K**-0.5` |
+| `chunk_size` | `64` | `64/128` |
+| `output_final_state` | `false` | bool |
+| `safe_gate` | `false` | bool |
+| `lower_bound` | `-5.0` | safe raw gate 时 `[-5,0)` |
+| `use_gate_in_kernel` | `false` | bool |
+| `disable_recompute` | `false` | bool |
+| `return_intermediate_states` | `false` | bool |
+| `state_v_first` | `false` | bool |
+
+## 支持范围
+
+- A2 (`ascend910b`)、A3 (`ascend910_93`)、A5 (`ascend950`)。
+- `K/V` 为 `[16,256]` 内 16 的倍数；交付重点覆盖 K=128、V=128/256。
+- `chunk_size` 为 64/128。
+- TND/NTD 均支持多 head。
+- 变长调用最多 1024 条逻辑序列，rank-4 变长输入要求 B=1。
+
+## 验证
+
+唯一用例规格是 `tests/op_cases/chunk_kda_fwd.json`。数值测试位于
+`tests/operators/chunk_kda_fwd/accuracy/`，性能使用 `tests/operators/chunk_kda_fwd/performance/profile.py`
+和 `msopprof`。
+
+完整 API 见 [API 文档](docs/api.md)，阶段和内存设计见 [设计文档](docs/design.md)。