From e3cae9682eeff2d62af3157488094d8b6d1cf14f Mon Sep 17 00:00:00 2001 From: Stefan Koncarevic Date: Fri, 10 Jul 2026 16:04:06 +0000 Subject: [PATCH] [ROCm] [CI] Keep original layer count for Eagle3DeepseekV2 init test Eagle3 draft `AQ-MedAI/Kimi-K25-eagle3` requests auxiliary hidden states from target layers (2, 30, 58). The initialization test's default `dummy_hf_overrides` collapses the target `Kimi-K2.5` to a single hidden layer, so those layers don't exist and `DeepseekV2Model.forward` returns a bare `hidden_states` tensor instead of `(hidden_states, aux_hidden_states)`. With Eagle3's `use_aux_hidden_state_outputs=True`, `_dummy_run` unpacks two values and iterating the bare tensor fails: ValueError: too many values to unpack (expected 2) Set `use_original_num_layers=True` so the target keeps all layers, the configured aux layers exist, and the tuple is returned (same pattern as the other Eagle3 entries). Also set `max_model_len=8192`: with the full layer count Kimi-K2.5 is large, and the default max len sizes a KV cache too big to fit alongside the weights (OOM); reducing it keeps the test within CI VRAM. Test-only change; no vLLM runtime code is modified. Signed-off-by: Stefan Koncarevic --- tests/models/registry.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tests/models/registry.py b/tests/models/registry.py index fccfb549f505..deb833b8beab 100644 --- a/tests/models/registry.py +++ b/tests/models/registry.py @@ -1461,6 +1461,8 @@ def check_available_online( trust_remote_code=True, speculative_model="AQ-MedAI/Kimi-K25-eagle3", tokenizer="moonshotai/Kimi-K2.5", + use_original_num_layers=True, + max_model_len=8192, # Reduce KV cache so full-layer Kimi fits in CI VRAM ), "Eagle3DeepseekV3ForCausalLM": _HfExamplesInfo( "moonshotai/Kimi-K2.5",