[Model] Use upstream GLM-5.3 and Qwen4-Exp configs and processor (#57387)

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:
Harry Mellor
2026-10-02 02:25:18 +00:00
committed by GitHub
co-authored by Claude Opus 5
parent 866f02db8a
commit e3cae8d2ac
36 changed files with 366 additions and 2385 deletions
+12 -1
View File
@@ -108,6 +108,16 @@ HfOverrides = dict[str, Any] | Callable[[PreTrainedConfig], PreTrainedConfig]
ModelImpl = Literal["auto", "vllm", "transformers", "terratorch"]
LayerBlockType = Literal["attention", "linear_attention", "mamba"]
_ATTENTION_LAYER_TYPES = frozenset(
{
"full_attention",
"deepseek_sparse_attention",
"qwen_sparse_attention",
}
)
"""`layer_types` spellings that consume a full attention KV cache. Sparse
attention still caches every token, so it counts as attention here."""
_RUNNER_CONVERTS: dict[RunnerType, list[ConvertType]] = {
"generate": [],
"pooling": ["embed", "classify"],
@@ -1718,7 +1728,8 @@ class ModelConfig:
if layer_types_value is not None:
if block_type == "attention":
return sum(
t == "full_attention" for t in layer_types_value[start:end]
t in _ATTENTION_LAYER_TYPES
for t in layer_types_value[start:end]
)
elif block_type == "linear_attention":
return sum(