[Bugfix] Add Kimi K3 MoE support to benchmark_moe.py (#50082)
Signed-off-by: Vansh Bhatia <210711135+vanshbhatia-amd@users.noreply.github.com> Co-authored-by: Vansh Bhatia <210711135+vanshbhatia-amd@users.noreply.github.com>
This commit is contained in:
co-authored by
Vansh Bhatia
parent
541c6d64c1
commit
d591d1d511
@@ -813,6 +813,19 @@ def get_model_params(config):
|
|||||||
topk = config.thinker_config.text_config.num_experts_per_tok
|
topk = config.thinker_config.text_config.num_experts_per_tok
|
||||||
intermediate_size = config.thinker_config.text_config.moe_intermediate_size
|
intermediate_size = config.thinker_config.text_config.moe_intermediate_size
|
||||||
hidden_size = config.thinker_config.text_config.hidden_size
|
hidden_size = config.thinker_config.text_config.hidden_size
|
||||||
|
elif architecture in (
|
||||||
|
"KimiK3ForConditionalGeneration",
|
||||||
|
"KimiLinearForCausalLM",
|
||||||
|
):
|
||||||
|
# Kimi K3 (multimodal) nests its MoE params in a KimiLinearConfig
|
||||||
|
# text_config and uses ``num_experts_per_token`` rather than the more
|
||||||
|
# common ``num_experts_per_tok``. get_text_config() returns the config
|
||||||
|
# itself for the text-only KimiLinearForCausalLM.
|
||||||
|
text_config = config.get_text_config()
|
||||||
|
E = text_config.num_experts
|
||||||
|
topk = text_config.num_experts_per_token
|
||||||
|
intermediate_size = text_config.moe_intermediate_size
|
||||||
|
hidden_size = text_config.hidden_size
|
||||||
elif architecture == "PixtralForConditionalGeneration":
|
elif architecture == "PixtralForConditionalGeneration":
|
||||||
# Pixtral can contain different LLM architectures,
|
# Pixtral can contain different LLM architectures,
|
||||||
# recurse to get their parameters
|
# recurse to get their parameters
|
||||||
|
|||||||
Reference in New Issue
Block a user