[Ops][Feature] Add Ascend 310P support for chunk_gated_delta_rule_fwd_h and chunk_fwd_o (vllm-project#9458)

## Summary
- Ports chunk_gated_delta_rule_fwd_h (gated delta rule forward for
DeltaNet/Qwen3) from 910B to 310P
- Ports chunk_fwd_o to 310p
- All 310P changes gated by #ifdef CATLASS_UNIFIED_CORE — 910B code
unchanged
- Unified-core execution path (sequential cube→vector, no AIC/AIV split)
- L0C→UB→GM bridge: BLOCK_MODE_MATRIX + fractal-by-fractal strided
DataCopy (no Fixpipe on 310P)
- CAST_RINT→CAST_NONE for float→half (CAST_RINT silently produces zeros
on 310P)
- DataCopyPad→DataCopy (DataCopyPad unsupported on 310P)
- compat_310p.h: dummy bfloat16_t (310P only), PIPE_FIX→PIPE_MTE3,
LoadDataWithSparse shim
- Preserves original 8 dtype combos (bf16+fp16) for 910B/910C
compatibility

## Test plan
- [x] Precision test via ctypes (test_aclnn_ctypes.py): all outputs
cos>=0.999998 on 310P3
- [x] Python API test (test_npu_chunk_gdr.py): 4 shapes PASS
- [x] ruff check: PASS
- [x] format.sh ci: PASS

<img width="1463" height="541" alt="Screenshot 2026-05-25 at 11 52 28"
src="https://github.com/user-attachments/assets/d56838c7-e81b-4123-afcb-861599c08dae"
/>


- vLLM version: v0.22.1
- vLLM main:
vllm-project/vllm@967c5c3

---------

Signed-off-by: Vladimir Evmenov <evmenoff@Vladimirs-MacBook-Air.local>
Signed-off-by: Vladimir Evmenov <vladimir.evmenov.jr@gmail.com>
Signed-off-by: Tflowers-0129 <2906339855@qq.com>
Co-authored-by: Vladimir Evmenov <evmenoff@Vladimirs-MacBook-Air.local>
Co-authored-by: Tflowers-0129 <2906339855@qq.com>

Author: 3 people

.github/workflows/_e2e_nightly_single_node.yaml

@@ -124,6 +124,7 @@ jobs:
           ref: ${{ inputs.vllm_ascend_ref }}
           path: ./temp-vllm-ascend
           fetch-depth: 1
+          submodules: recursive
 
       - name: Move code to /vllm-workspace
         if: ${{ inputs.request_id != '' }}

.github/workflows/_e2e_nightly_single_node_models.yaml

@@ -112,6 +112,7 @@ jobs:
           ref: ${{ inputs.vllm_ascend_ref }}
           path: ./temp-vllm-ascend
           fetch-depth: 1
+          submodules: recursive
 
       - name: Move vllm-ascend code to /vllm-workspace
         if: ${{ inputs.request_id != '' }}

.github/workflows/_schedule_image_build.yaml

@@ -140,20 +140,23 @@ jobs:
     - uses: actions/checkout@v7
       if: ${{ github.event_name != 'workflow_dispatch' }}
       with:
+        submodules: recursive
         fetch-depth: 0
         persist-credentials: false
         ref: ${{ inputs.branch_ref }}
 
     - uses: actions/checkout@v7
       if: ${{ github.event_name == 'workflow_dispatch' && inputs.workflow_dispatch_tag != '' }}
       with:
+        submodules: recursive
         fetch-depth: 0
         persist-credentials: false
         ref: ${{ inputs.workflow_dispatch_tag }}
 
     - uses: actions/checkout@v7
       if: ${{ github.event_name == 'workflow_dispatch' && inputs.workflow_dispatch_tag == '' && inputs.vllm_ascend_commit == '' }}
       with:
+        submodules: recursive
         fetch-depth: 0
         persist-credentials: false
         ref: ${{ inputs.branch_ref }}
@@ -299,12 +302,14 @@ jobs:
         uses: actions/checkout@v7
         if: ${{ github.event_name != 'workflow_dispatch' || (github.event_name == 'workflow_dispatch' && inputs.workflow_dispatch_tag == '') }}
         with:
+          submodules: recursive
           ref: ${{ inputs.branch_ref }}
 
       - name: Checkout tag
         uses: actions/checkout@v7
         if: ${{ github.event_name == 'workflow_dispatch' && inputs.workflow_dispatch_tag != '' }}
         with:
+          submodules: recursive
           ref: ${{ inputs.workflow_dispatch_tag }}
 
       - name: Download arm64 digests

.github/workflows/_selected_tests.yaml

@@ -110,6 +110,7 @@ jobs:
         with:
           ref: ${{ inputs.ref || github.ref }}
           fetch-depth: 0
+          submodules: recursive
 
       - name: Rebase on latest main
         run: |
@@ -203,6 +204,22 @@ jobs:
             vllm_ascend/include
           key: vllm-ascend-build-v1-${{ steps.get_arch.outputs.arch }}-${{ matrix.group.image_tag }}-${{ steps.get_csrc_hash.outputs.CSRC_HASH }}
 
+      - name: Verify required AscendC custom ops are registered (310p)
+        if: ${{ matrix.group.npu_type == '310p' }}
+        run: |
+          python - <<'PY'
+          import sys, torch
+          import vllm_ascend.vllm_ascend_C  # noqa: F401
+          ops = torch.ops._C_ascend
+          missing = [n for n in ("chunk_gated_delta_rule_fwd_h", "chunk_fwd_o") if not hasattr(ops, n)]
+          if missing:
+              print(f"::error::Missing AscendC ops after install: {missing}. "
+                    f"Cache contents are stale or build skipped registration.", file=sys.stderr)
+              sys.exit(1)
+          print("AscendC chunk-gdr ops registered:",
+                [n for n in ("chunk_gated_delta_rule_fwd_h", "chunk_fwd_o") if hasattr(ops, n)])
+          PY
+
       - name: Install vllm-project/vllm-ascend no device
         if: ${{ matrix.group.npu_type == 'cpu' }}
         env:

.github/workflows/push_build_csrc_cache.yaml

@@ -102,6 +102,8 @@ jobs:
     steps:
       - name: Checkout repo
         uses: actions/checkout@v7
+        with:
+          submodules: recursive
 
       - name: Config mirrors (ubuntu)
         if: matrix.os == 'ubuntu'

.github/workflows/schedule_release_code_and_wheel.yml

@@ -49,11 +49,14 @@ jobs:
       - name: checkout vllm-ascend
         if: ${{ github.event_name != 'workflow_dispatch' }}
         uses: actions/checkout@v7
+        with:
+          submodules: recursive
 
       - name: checkout vllm-ascend ${{ inputs.tag }}
         if: ${{ github.event_name == 'workflow_dispatch' }}
         uses: actions/checkout@v7
         with:
+          submodules: recursive
           ref: ${{ inputs.tag }}
 
       - name: Print
@@ -99,11 +102,14 @@ jobs:
       - name: checkout vllm-ascend
         if: ${{ github.event_name != 'workflow_dispatch' }}
         uses: actions/checkout@v7
+        with:
+          submodules: recursive
 
       - name: checkout vllm-ascend ${{ inputs.tag }}
         if: ${{ github.event_name == 'workflow_dispatch' }}
         uses: actions/checkout@v7
         with:
+          submodules: recursive
           ref: ${{ inputs.tag }}
 
       - name: Free up disk space
@@ -183,11 +189,14 @@ jobs:
       - name: checkout vllm-ascend
         if: ${{ github.event_name != 'workflow_dispatch' }}
         uses: actions/checkout@v7
+        with:
+          submodules: recursive
 
       - name: checkout vllm-ascend ${{ inputs.tag }}
         if: ${{ github.event_name == 'workflow_dispatch' }}
         uses: actions/checkout@v7
         with:
+          submodules: recursive
           ref: ${{ inputs.tag }}
 
       - name: Free up disk space
@@ -261,11 +270,14 @@ jobs:
       - name: checkout vllm-ascend
         if: ${{ github.event_name != 'workflow_dispatch' }}
         uses: actions/checkout@v7
+        with:
+          submodules: recursive
 
       - name: checkout vllm-ascend ${{ inputs.tag }}
         if: ${{ github.event_name == 'workflow_dispatch' }}
         uses: actions/checkout@v7
         with:
+          submodules: recursive
           ref: ${{ inputs.tag }}
 
       - name: Free up disk space

csrc/build_aclnn.sh

@@ -57,6 +57,8 @@ if [[ "$SOC_VERSION" =~ ^ascend310 ]]; then
     CUSTOM_OPS_ARRAY=(
         "causal_conv1d_v310"
         "recurrent_gated_delta_rule_v310"
+        "chunk_fwd_o"
+        "chunk_gated_delta_rule_fwd_h"
     )
     CUSTOM_OPS=$(IFS=';'; echo "${CUSTOM_OPS_ARRAY[*]}")
     SOC_ARG="ascend310p"

csrc/moe/chunk_fwd_o/op_host/chunk_fwd_o_def.cpp

@@ -96,6 +96,7 @@ class ChunkFwdO : public OpDef {
         this->AICore().AddConfig("ascend910b", aicore_config);
         this->AICore().AddConfig("ascend910_93", aicore_config);
         this->AICore().AddConfig("ascend950", aicore_config);
+        this->AICore().AddConfig("ascend310p", aicore_config);
     }
 };
 

csrc/moe/chunk_fwd_o/op_kernel/arch20/compat_310p.h

@@ -0,0 +1,39 @@
+#ifndef COMPAT_310P_H
+#define COMPAT_310P_H
+
+#ifndef __CCE_KT_TEST__
+#include "kernel_operator.h"
+#endif
+
+// Dummy bfloat16_t only needed on 310P (dav_m200) where the compiler
+// doesn't provide a native bf16 type. On 910B/910C the compiler's
+// __clang_cce_types.h already typedefs bfloat16_t from __bf16.
+#if defined(__CCE_AICORE__) && (__CCE_AICORE__ == 200) && !defined(__bfloat16_t_defined)
+#define __bfloat16_t_defined
+#define __COMPAT_310P_ACTIVE__
+struct bfloat16_t {
+    uint16_t val;
+    bfloat16_t() = default;
+    bfloat16_t(float v) : val(0) { (void)v; }
+    operator float() const { return 0.f; }
+};
+#endif
+
+// 310P has no fixpipe unit; post-matmul stores go through MTE3
+#ifndef PIPE_FIX
+#define PIPE_FIX PIPE_MTE3
+#endif
+
+// 310P renames LoadDataWithSparse → LoadDataWithSparseCal
+#ifdef __COMPAT_310P_ACTIVE__
+#define LoadDataWithSparse LoadDataWithSparseCal
+#endif
+
+// 310P has no AscendC::ToFloat — dummy bfloat16_t already has operator float()
+#ifdef __COMPAT_310P_ACTIVE__
+namespace AscendC {
+    inline float ToFloat(bfloat16_t v) { return (float)v; }
+}
+#endif
+
+#endif