[BugFix][Model] Fix the GLM-5.3-Flash NoPE and KDA convolution paths on Ascend

tests/ut/attention/test_mla_nope_cache_paths.py

@@ -0,0 +1,96 @@
+# SPDX-License-Identifier: Apache-2.0
+"""MLA-NoPE layers must keep their own KV cache write path.
+
+Two notions of "no RoPE" meet in ``AscendMLAImpl``. ``use_mla_rope`` is false
+whenever the layer has no rotary embedding, which for GLM-5.3-Flash follows
+``mla_nope``; ``qk_rope_head_dim == 0`` additionally says the rope *cache* is
+empty. Kimi K3 is the first case without the second, and its ``_exec_kv_no_rope``
+writes both caches through ``reshape_and_cache``. GLM-5.3-Flash is both, so it
+needs ``_exec_kv_mla_nope``, which scatters only the nope cache and tolerates a
+padded (non-contiguous) one.
+
+Guarding the K3 path on ``use_mla_rope`` alone shadowed the NoPE branches, and
+letting the fused decode prolog run on a NoPE layer fed the operator an empty
+rope cache -- whose torch strides it rejects, since contiguous strides are
+computed over ``max(size, 1)`` and so report 1 where the tiling wants 0:
+
+    krCache dim2 must be contiguous, actual stride is 1, expected stride is 0
+
+Both live inside methods that need a device to construct, hence the
+source-level checks.
+"""
+
+from __future__ import annotations
+
+import ast
+from pathlib import Path
+
+ROOT = Path(__file__).resolve().parents[3]
+MLA_V1 = ROOT / "vllm_ascend" / "attention" / "mla_v1.py"
+
+ROPE_CACHE_EXISTS = "self.qk_rope_head_dim > 0"
+K3_GUARD = f"not self.use_mla_rope and {ROPE_CACHE_EXISTS}"
+
+
+def _function(name: str) -> ast.FunctionDef:
+    tree = ast.parse(MLA_V1.read_text(encoding="utf-8"))
+    for node in ast.walk(tree):
+        if isinstance(node, ast.FunctionDef) and node.name == name:
+            return node
+    raise AssertionError(f"{name} not found in {MLA_V1}")
+
+
+def _calls(node: ast.AST) -> set[str]:
+    return {ast.unparse(inner.func) for inner in ast.walk(node) if isinstance(inner, ast.Call)}
+
+
+def _decode_prolog_gate() -> str:
+    for node in ast.walk(_function("forward")):
+        if isinstance(node, ast.Assign) and ast.unparse(node.targets[0]) == "can_use_decode_prolog":
+            return ast.unparse(node.value)
+    raise AssertionError("can_use_decode_prolog is no longer assigned in forward")
+
+
+def test_decode_prolog_needs_a_rope_cache() -> None:
+    gate = _decode_prolog_gate()
+
+    assert ROPE_CACHE_EXISTS in gate, (
+        "the fused decode prolog is enabled without checking that a rope cache "
+        f"exists; NoPE layers will hit the tiling error. Gate: {gate}"
+    )
+    # Hardware support is necessary but not sufficient.
+    assert "MLA_DECODE_PROLOG_WITHOUT_ROPE" in gate
+    assert "self.use_mla_rope" in gate
+
+
+def test_decode_prolog_still_allowed_with_rope() -> None:
+    """Layers that do have RoPE must keep the fused path."""
+    gate = _decode_prolog_gate()
+
+    assert gate.startswith("self.use_mla_rope or")
+
+
+def _k3_guards(function_name: str) -> list[str]:
+    return [
+        ast.unparse(node.test)
+        for node in ast.walk(_function(function_name))
+        if isinstance(node, ast.If) and "use_mla_rope" in ast.unparse(node.test)
+    ]
+
+
+def test_kimi_k3_path_is_gated_on_a_rope_cache() -> None:
+    for function_name in ("exec_kv_decode", "exec_kv_prefill"):
+        guards = _k3_guards(function_name)
+
+        assert K3_GUARD in guards, (
+            f"{function_name} routes every rope-less layer to _exec_kv_no_rope; "
+            f"NoPE layers never reach _exec_kv_mla_nope. Guards: {guards}"
+        )
+
+
+def test_nope_branch_is_reachable_from_both_phases() -> None:
+    """The NoPE scatter is the point of the guard above; keep both call sites."""
+    for function_name in ("exec_kv_decode", "exec_kv_prefill"):
+        assert "self._exec_kv_mla_nope" in _calls(_function(function_name)), (
+            f"{function_name} no longer calls _exec_kv_mla_nope"
+        )

tests/ut/compilation/test_qknorm_rope_fusion_pass.py

@@ -0,0 +1,76 @@
+#
+# 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.
+#
+from unittest.mock import MagicMock, patch
+
+import torch
+from vllm.config import VllmConfig
+
+from tests.ut.base import TestBase
+from vllm_ascend.compilation.passes.qknorm_rope_fusion_pass import QKNormRopeFusionPass
+
+PASS_MODULE = "vllm_ascend.compilation.passes.qknorm_rope_fusion_pass"
+
+
+def _vllm_config(dtype: torch.dtype = torch.bfloat16) -> VllmConfig:
+    vllm_config = VllmConfig()
+    vllm_config.model_config = MagicMock(dtype=dtype)
+    return vllm_config
+
+
+class TestQKNormRopeFusionPassGating(TestBase):
+    """The pass must not build patterns it cannot trace.
+
+    Registering a pattern allocates NPU example inputs and traces the rope
+    kernel, so these tests assert on whether the pattern classes get
+    instantiated rather than on the registered patterns themselves.
+    """
+
+    def setUp(self):
+        layer = MagicMock(head_size=128, num_heads=32, num_kv_heads=8)
+        patcher = patch(f"{PASS_MODULE}.get_layers_from_vllm_config", return_value={"layer": layer})
+        patcher.start()
+        self.addCleanup(patcher.stop)
+
+    @patch(f"{PASS_MODULE}.QKNormRopeFusionPatternWithBias")
+    @patch(f"{PASS_MODULE}.QKNormRopeFusionPattern")
+    @patch(f"{PASS_MODULE}.get_rope_dim", return_value=0)
+    def test_model_without_rope_registers_nothing(self, _rope_dim, pattern, pattern_with_bias):
+        # GLM-5.3-Flash reports qk_rope_head_dim == 0, so tracing the pattern
+        # would ask Triton for a zero-width arange.
+        QKNormRopeFusionPass(_vllm_config())
+
+        pattern.assert_not_called()
+        pattern_with_bias.assert_not_called()
+
+    @patch(f"{PASS_MODULE}.QKNormRopeFusionPatternWithBias")
+    @patch(f"{PASS_MODULE}.QKNormRopeFusionPattern")
+    @patch(f"{PASS_MODULE}.get_rope_dim", return_value=128)
+    def test_model_with_rope_registers_both_patterns(self, _rope_dim, pattern, pattern_with_bias):
+        QKNormRopeFusionPass(_vllm_config())
+
+        # One registration per supported epsilon.
+        self.assertEqual(pattern.call_count, 2)
+        self.assertEqual(pattern_with_bias.call_count, 2)
+
+    @patch(f"{PASS_MODULE}.QKNormRopeFusionPatternWithBias")
+    @patch(f"{PASS_MODULE}.QKNormRopeFusionPattern")
+    @patch(f"{PASS_MODULE}.get_rope_dim", return_value=128)
+    def test_unsupported_dtype_registers_nothing(self, _rope_dim, pattern, pattern_with_bias):
+        QKNormRopeFusionPass(_vllm_config(dtype=torch.float16))
+
+        pattern.assert_not_called()
+        pattern_with_bias.assert_not_called()