llama: dynamic head_dim and n_rot for SWA (#20301)

* llama: dynamic head_dim and n_rot for SWA

* also add gguf_writer wrappers

* fix build

* build_rope_shift arg reorder
This commit is contained in:
Xuan-Son Nguyen
2026-03-09 22:22:39 +01:00
committed by GitHub
parent 23fbfcb1ad
commit 59db9a357d
112 changed files with 419 additions and 346 deletions
+9
View File
@@ -773,6 +773,12 @@ class GGUFWriter:
def add_value_length_mla(self, length: int) -> None:
self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length)
def add_key_length_swa(self, length: int) -> None:
self.add_uint32(Keys.Attention.KEY_LENGTH_SWA.format(arch=self.arch), length)
def add_value_length_swa(self, length: int) -> None:
self.add_uint32(Keys.Attention.VALUE_LENGTH_SWA.format(arch=self.arch), length)
def add_indexer_head_count(self, count: int) -> None:
self.add_uint32(Keys.Attention.Indexer.HEAD_COUNT.format(arch=self.arch), count)
@@ -946,6 +952,9 @@ class GGUFWriter:
def add_rope_dimension_count(self, count: int) -> None:
self.add_uint32(Keys.Rope.DIMENSION_COUNT.format(arch=self.arch), count)
def add_rope_dimension_count_swa(self, count: int) -> None:
self.add_uint32(Keys.Rope.DIMENSION_COUNT_SWA.format(arch=self.arch), count)
def add_rope_dimension_sections(self, dims: Sequence[int]) -> None:
self.add_array(Keys.Rope.DIMENSION_SECTIONS.format(arch=self.arch), dims)