From 97cc8729f0bc351a5536380fd897607f4ecdeef1 Mon Sep 17 00:00:00 2001
From: Ekagra Ranjan <3116519+ekagra-ranjan@users.noreply.github.com>
Date: Tue, 29 Apr 2025 03:30:40 -0400
Subject: [PATCH] [Model] Ignore rotary embed load for Cohere model (#17319)

---
 vllm/model_executor/models/commandr.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/vllm/model_executor/models/commandr.py b/vllm/model_executor/models/commandr.py
index 25b1d5a195..8f64e5d5c9 100644
--- a/vllm/model_executor/models/commandr.py
+++ b/vllm/model_executor/models/commandr.py
@@ -418,6 +418,10 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
         loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
 
+            # Skip loading rotary embeddings since vLLM has its own
+            if "rotary_emb.inv_freq" in name:
+                continue
+
             if (self.quant_config is not None and
                 (scale_name := self.quant_config.get_cache_scale(name))):
                 # Loading kv cache quantization scales