Add Mistral Large 3 and Ministral 3 (#29757)

Signed-off-by: Julien Denize <julien.denize@mistral.ai>
Signed-off-by: Julien Denize <40604584+juliendenize@users.noreply.github.com>
Signed-off-by: Mickael Seznec <mickael@mistral.ai>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Roger Wang <hey@rogerw.io>
Co-authored-by: Mickael Seznec <mickael@mistral.ai>
This commit is contained in:
Julien Denize
2025-12-02 10:29:00 +00:00
committed by GitHub
co-authored by Roger Wang Mickael Seznec
parent 8bbcf8b6e7
commit d8c6210eea
16 changed files with 724 additions and 30 deletions
+4
View File
@@ -111,6 +111,7 @@ class MultiHeadLatentAttentionWrapper(CustomOp):
self,
positions: torch.Tensor,
hidden_states: torch.Tensor,
llama_4_scaling: torch.Tensor | None = None,
) -> torch.Tensor:
q_c = None
kv_lora = None
@@ -159,6 +160,9 @@ class MultiHeadLatentAttentionWrapper(CustomOp):
hidden_states, q_c, positions, self.indexer_rope_emb
)
if llama_4_scaling is not None:
q *= llama_4_scaling
attn_out = self.mla_attn(
q,
kv_c_normed,