model : add sarvam_moe architecture support (#20275)

This commit is contained in:
Sumit Chatterjee
2026-05-10 00:31:50 +10:00
committed by GitHub
parent 65d7a8bbf0
commit 1e5ad35d56
4 changed files with 46 additions and 0 deletions
+31
View File
@@ -1570,6 +1570,9 @@ class TextModel(ModelBase):
if chkhsh == "862f827721df956049dff5ca81a57f29e575280bc622e290d3bf4e35eca29015":
# ref: https://huggingface.co/codefuse-ai/F2LLM-v2-4B
res = "f2llmv2"
if chkhsh == "62f6fb0a6fd5098caeabb19b07a5c1099cafc8b9c40eab6ea89ece4ec02fbc57":
# ref: https://huggingface.co/sarvamai/sarvam-30b
res = "sarvam-moe"
if res is None:
logger.warning("\n")
@@ -11591,6 +11594,34 @@ class BailingMoeV2Model(TextModel):
raise ValueError(f"Unprocessed experts: {experts}")
@ModelBase.register("SarvamMoEForCausalLM", "modeling_sarvam_moe.SarvamMoEForCausalLM")
class SarvamMoEModel(BailingMoeV2Model):
model_arch = gguf.MODEL_ARCH.BAILINGMOE2
# Sarvam-MoE shares the BailingMoeV2 architecture; only differences:
# - full rotary (no partial_rotary_factor)
# - expert bias is zero-mean normalized at load time
def set_gguf_parameters(self):
super().set_gguf_parameters()
hparams = self.hparams
if (rope_dim := hparams.get("head_dim")) is None:
rope_dim = hparams["hidden_size"] // hparams["num_attention_heads"]
# Override the partial-rotary value written by BailingMoeV2 with the full rotary dim
self.gguf_writer.add_rope_dimension_count(rope_dim)
@classmethod
def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
name, gen = item
if name.endswith(".expert_bias"):
# Sarvam normalizes expert bias to zero mean
inner = gen
def gen():
t = inner()
return t - t.mean()
return super().filter_tensors((name, gen))
@ModelBase.register("GroveMoeForCausalLM", "modeling_grove_moe.GroveMoeForCausalLM")
class GroveMoeModel(TextModel):
model_arch = gguf.MODEL_ARCH.GROVEMOE