Make `config.json` compatible with standard sliding window config
Browse filesThis will add `layer_types` to the loaded config class so that libraries such as vLLM can load hybrid attention models in the standard Hugging Face format.
Since we do not edit `configuration_phi4flash.py` this change is backwards compatible.
Once this change has been merged along with https://github.com/vllm-project/vllm/pull/21927 we can update `configuration_phi4flash.py` so that the the modelling code works in the standard way too.
- config.json +6 -0
config.json
CHANGED
@@ -26,6 +26,12 @@
|
|
26 |
"num_key_value_heads": 20,
|
27 |
"resid_pdrop": 0.0,
|
28 |
"sliding_window": 512,
|
|
|
|
|
|
|
|
|
|
|
|
|
29 |
"torch_dtype": "bfloat16",
|
30 |
"tie_word_embeddings": true,
|
31 |
"transformers_version": "4.46.1",
|
|
|
26 |
"num_key_value_heads": 20,
|
27 |
"resid_pdrop": 0.0,
|
28 |
"sliding_window": 512,
|
29 |
+
"layer_types": [
|
30 |
+
"full_attention", "sliding_attention", "full_attention", "sliding_attention", "full_attention", "sliding_attention", "full_attention", "sliding_attention",
|
31 |
+
"full_attention", "sliding_attention", "full_attention", "sliding_attention", "full_attention", "sliding_attention", "full_attention", "sliding_attention",
|
32 |
+
"full_attention", "full_attention", "full_attention", "full_attention", "full_attention", "full_attention", "full_attention", "full_attention",
|
33 |
+
"full_attention", "full_attention", "full_attention", "full_attention", "full_attention", "full_attention", "full_attention", "full_attention"
|
34 |
+
],
|
35 |
"torch_dtype": "bfloat16",
|
36 |
"tie_word_embeddings": true,
|
37 |
"transformers_version": "4.46.1",
|