[MoE] Add LFM2 MoE tuning support + tuned configs for H100/B200/MI325X (#22791)

Co-authored-by: Piotr Mazurek <piotr.mazurek@liquid.ai>
This commit is contained in:
Piotr Mazurek
2026-04-21 18:32:05 -07:00
committed by GitHub
co-authored by Piotr Mazurek
parent 0e165ffbfc
commit 6cf0b004ca
25 changed files with 3852 additions and 0 deletions
@@ -138,6 +138,10 @@ def get_model_config(
E = config.num_experts // ep_size
topk = config.top_k_experts
intermediate_size = config.moe_intermediate_size
elif architecture == "Lfm2MoeForCausalLM":
E = config.num_experts // ep_size
topk = config.num_experts_per_tok
intermediate_size = config.moe_intermediate_size
else:
# Default: Mixtral
E = config.num_local_experts // ep_size