CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-07 09:57:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

fixes

Browse Source
This commit is contained in:
younesbelkada
2025-07-04 14:47:31 +04:00
parent 14c37ec047
commit 50eadc7b33
6 changed files with 28 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
convert_hf_to_gguf.py
View File

@@ -6632,11 +6632,11 @@ class FalconH1Model(Mamba2Model):
self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_key_length(self.hparams["head_dim"])
self.gguf_writer.add_value_length(self.hparams["head_dim"])
self.gguf_writer.add_float32("falcon-h1.key_multiplier", self.hparams["key_multiplier"])
self.gguf_writer.add_float32("falcon_h1.key_multiplier", self.hparams["key_multiplier"])
## Other params
self.gguf_writer.add_float32("falcon-h1.lm_head_multiplier", self.hparams["lm_head_multiplier"])
self.gguf_writer.add_float32("falcon-h1.embedding_multiplier", self.hparams["embedding_multiplier"])
self.gguf_writer.add_float32("falcon_h1.lm_head_multiplier", self.hparams["lm_head_multiplier"])
self.gguf_writer.add_float32("falcon_h1.embedding_multiplier", self.hparams["embedding_multiplier"])
## Validation ##
assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported"
@@ -6644,34 +6644,34 @@ class FalconH1Model(Mamba2Model):
# Add Falcon Mamba2 specific configuration
self.gguf_writer.add_uint32("falcon-h1.ssm.mamba_chunk_size", self.hparams["mamba_chunk_size"])
self.gguf_writer.add_uint32("falcon-h1.attention.head_dim", self.hparams["head_dim"])
self.gguf_writer.add_uint32("falcon-h1.ssm.mamba_d_ssm", self.hparams["mamba_d_ssm"])
self.gguf_writer.add_uint32("falcon-h1.num_attention_heads", self.find_hparam(["num_attention_heads"]))
self.gguf_writer.add_uint32("falcon-h1.num_key_value_heads",
self.gguf_writer.add_uint32("falcon_h1.ssm.mamba_chunk_size", self.hparams["mamba_chunk_size"])
self.gguf_writer.add_uint32("falcon_h1.attention.head_dim", self.hparams["head_dim"])
self.gguf_writer.add_uint32("falcon_h1.ssm.mamba_d_ssm", self.hparams["mamba_d_ssm"])
self.gguf_writer.add_uint32("falcon_h1.num_attention_heads", self.find_hparam(["num_attention_heads"]))
self.gguf_writer.add_uint32("falcon_h1.num_key_value_heads",
self.find_hparam(["num_key_value_heads"], optional=True) or
self.find_hparam(["num_attention_heads"]))
# Add multipliers as metadata instead of tensors
self.gguf_writer.add_float32("falcon-h1.attention_in_multiplier", self.attention_in_multiplier)
self.gguf_writer.add_float32("falcon-h1.attention_out_multiplier", self.attention_out_multiplier)
self.gguf_writer.add_float32("falcon-h1.ssm_in_multiplier", self.ssm_in_multiplier)
self.gguf_writer.add_float32("falcon-h1.ssm_out_multiplier", self.ssm_out_multiplier)
self.gguf_writer.add_float32("falcon_h1.attention_in_multiplier", self.attention_in_multiplier)
self.gguf_writer.add_float32("falcon_h1.attention_out_multiplier", self.attention_out_multiplier)
self.gguf_writer.add_float32("falcon_h1.ssm_in_multiplier", self.ssm_in_multiplier)
self.gguf_writer.add_float32("falcon_h1.ssm_out_multiplier", self.ssm_out_multiplier)
# Add MLP multipliers
if isinstance(self.mlp_multipliers, (list, tuple)) and len(self.mlp_multipliers) == 2:
self.gguf_writer.add_float32("falcon-h1.mlp_gate_multiplier", self.mlp_multipliers[0])
self.gguf_writer.add_float32("falcon-h1.mlp_down_multiplier", self.mlp_multipliers[1])
self.gguf_writer.add_float32("falcon_h1.mlp_gate_multiplier", self.mlp_multipliers[0])
self.gguf_writer.add_float32("falcon_h1.mlp_down_multiplier", self.mlp_multipliers[1])
# Add has MuP flag if SSM multipliers are present
if self.ssm_multipliers is not None:
self.gguf_writer.add_bool("falcon-h1.ssm.has_mup", True)
self.gguf_writer.add_bool("falcon_h1.ssm.has_mup", True)
# Add any other Falcon Mamba2 specific configuration
self.gguf_writer.add_bool("falcon-h1.mamba_use_mlp", self.find_hparam(["mamba_use_mlp"], optional=True))
self.gguf_writer.add_bool("falcon-h1.mamba_norm_before_gate", self.find_hparam(["mamba_norm_before_gate"], optional=True))
self.gguf_writer.add_bool("falcon-h1.mamba_rms_norm", self.find_hparam(["mamba_rms_norm"], optional=True))
self.gguf_writer.add_float32("falcon-h1.rope_theta", self.find_hparam(["rope_theta"], optional=True))
self.gguf_writer.add_bool("falcon_h1.mamba_use_mlp", self.find_hparam(["mamba_use_mlp"], optional=True))
self.gguf_writer.add_bool("falcon_h1.mamba_norm_before_gate", self.find_hparam(["mamba_norm_before_gate"], optional=True))
self.gguf_writer.add_bool("falcon_h1.mamba_rms_norm", self.find_hparam(["mamba_rms_norm"], optional=True))
self.gguf_writer.add_float32("falcon_h1.rope_theta", self.find_hparam(["rope_theta"], optional=True))
###### CONVERSION LOGIC ######