[Bugfix] Align vLLM YaRN with Transformers and stop re-scaling max_model_len (#56446)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
a5b714eee4
commit
c191787a68
@@ -6,6 +6,7 @@ only get the `eos_token_id` from the tokenizer as defined by
|
||||
`BaseRenderer.get_eos_token_id`.
|
||||
"""
|
||||
|
||||
import math
|
||||
from types import SimpleNamespace
|
||||
from typing import cast
|
||||
from unittest.mock import MagicMock, patch
|
||||
@@ -52,6 +53,31 @@ def test_patch_legacy_rope_type_preserves_nope_layers():
|
||||
}
|
||||
|
||||
|
||||
def test_patch_legacy_rope_type_normalizes_telechat3_yarn():
|
||||
"""TeleChat3's RoPE is YaRN with 0.07 in place of the usual 0.1.
|
||||
|
||||
Encoding that as a precomputed attention_factor keeps the config
|
||||
plain YaRN, which Transformers and every "yarn" guard understand.
|
||||
`mscale` cannot express it: Transformers only applies mscale when
|
||||
mscale_all_dim is also truthy.
|
||||
"""
|
||||
rope_parameters = {
|
||||
"type": "telechat3-yarn",
|
||||
"rope_type": "telechat3-yarn",
|
||||
"factor": 4.0,
|
||||
"original_max_position_embeddings": 8192,
|
||||
}
|
||||
|
||||
patch_legacy_rope_type(rope_parameters)
|
||||
|
||||
assert rope_parameters == {
|
||||
"rope_type": "yarn",
|
||||
"factor": 4.0,
|
||||
"original_max_position_embeddings": 8192,
|
||||
"attention_factor": pytest.approx(0.07 * math.log(4.0) + 1.0),
|
||||
}
|
||||
|
||||
|
||||
def test_glm5_next_accepts_deepseek_sparse_attention_layers():
|
||||
layer_types = ["linear_attention", "deepseek_sparse_attention"]
|
||||
|
||||
|
||||
Reference in New Issue
Block a user