From 4c56e62c85cea8fc2251efc25159836c214402aa Mon Sep 17 00:00:00 2001 From: "Kevin H. Luu" Date: Mon, 24 Aug 2026 08:16:28 -0700 Subject: [PATCH] [Bugfix][Kimi K3] Skip absent metadata during CUDA graph profiling (#53581) Signed-off-by: Kevin Luu <51931015+khluu@users.noreply.github.com> Co-authored-by: Codex --- tests/models/kimi_k3/test_kda.py | 14 ++++++++++++++ vllm/models/kimi_k3/amd/kda.py | 4 +++- vllm/models/kimi_k3/nvidia/kda.py | 4 +++- 3 files changed, 20 insertions(+), 2 deletions(-) diff --git a/tests/models/kimi_k3/test_kda.py b/tests/models/kimi_k3/test_kda.py index c8454e9e6f..3aba8aa3bc 100644 --- a/tests/models/kimi_k3/test_kda.py +++ b/tests/models/kimi_k3/test_kda.py @@ -20,6 +20,7 @@ from vllm.model_executor.layers.mamba.ops.gather_initial_states import ( from vllm.models.kimi_k3.amd.ops.third_party.kda import ( fused_recurrent_kda_packed_decode as fused_recurrent_kda_packed_decode_amd, ) +from vllm.models.kimi_k3.nvidia import kda as nvidia_kda from vllm.models.kimi_k3.nvidia.kda import ( _flashkda_prefill, _store_cache_checkpoints_kernel, @@ -59,6 +60,19 @@ PACKED_DECODE_IMPLS = { } +def test_kda_warmup_skips_missing_metadata(monkeypatch): + monkeypatch.setattr( + nvidia_kda, + "get_forward_context", + lambda: SimpleNamespace(attn_metadata={}), + ) + layer = object.__new__(nvidia_kda.KimiK3DeltaAttention) + object.__setattr__(layer, "prefix", "language_model.model.layers.0.self_attn") + empty = torch.empty(0, device=DEVICE) + + assert layer._forward(empty, empty, empty, empty, empty) is None + + def test_kda_recoverssm_config_state_layout(): vllm_config = SimpleNamespace( model_config=SimpleNamespace( diff --git a/vllm/models/kimi_k3/amd/kda.py b/vllm/models/kimi_k3/amd/kda.py index 870c13a3ed..e20a6136e1 100644 --- a/vllm/models/kimi_k3/amd/kda.py +++ b/vllm/models/kimi_k3/amd/kda.py @@ -313,7 +313,9 @@ class KimiK3DeltaAttention(GatedDeltaNetAttention): return assert isinstance(attn_metadata_raw, dict) - attn_metadata_narrowed = attn_metadata_raw[self.prefix] + attn_metadata_narrowed = attn_metadata_raw.get(self.prefix) + if attn_metadata_narrowed is None: + return assert isinstance(attn_metadata_narrowed, GDNAttentionMetadata) m = attn_metadata_narrowed has_initial_state = m.has_initial_state diff --git a/vllm/models/kimi_k3/nvidia/kda.py b/vllm/models/kimi_k3/nvidia/kda.py index d2ca6f42f5..9347751a25 100644 --- a/vllm/models/kimi_k3/nvidia/kda.py +++ b/vllm/models/kimi_k3/nvidia/kda.py @@ -658,7 +658,9 @@ class KimiK3DeltaAttention(GatedDeltaNetAttention): ) assert isinstance(attn_metadata_raw, dict) - attn_metadata_narrowed = attn_metadata_raw[self.prefix] + attn_metadata_narrowed = attn_metadata_raw.get(self.prefix) + if attn_metadata_narrowed is None: + return assert isinstance(attn_metadata_narrowed, KimiK3KDAMetadata) m = attn_metadata_narrowed has_initial_state = m.has_initial_state