Read max context length from the correct ModelArgs field

sxu · facebook-github-bot · commit 40002640eae5 · 2025-10-08T12:59:12.000-07:00
Summary: We should read from `max_context_len` for RoPE, `max_seq_len` represents input sequence length.

Differential Revision: D84182698
diff --git a/examples/models/llama/static_attention.py b/examples/models/llama/static_attention.py
@@ -259,7 +259,7 @@ def __init__(
         }
 
         rope = Rope(config)
-        freqs = rope.get_freqs(None, config.max_seq_len)
+        freqs = rope.get_freqs(None, config.max_context_len)
         self.freqs_cos = freqs[0].to(dtype)
         self.freqs_sin = freqs[1].to(dtype)
 

Original file line number	Diff line number	Diff line change
`@@ -259,7 +259,7 @@ def __init__(`
`259`	`259`	`}`
`260`	`260`
`261`	`261`	`rope = Rope(config)`
`262`		`- freqs = rope.get_freqs(None, config.max_seq_len)`
	`262`	`+ freqs = rope.get_freqs(None, config.max_context_len)`
`263`	`263`	`self.freqs_cos = freqs[0].to(dtype)`
`264`	`264`	`self.freqs_sin = freqs[1].to(dtype)`
`265`	`265`