happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'add-fh1-rebased' of https://github.com/tiiuae/llama.cpp-public into add-fh1-rebased

This commit is contained in:
ibrahimkhadraoui 2025-07-08 10:48:01 +04:00
parent 7846c67e5c 8555ee8b2c
commit f028a43a91
3 changed files with 7 additions and 11 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
convert_hf_to_gguf.py
View File

@ -6665,7 +6665,6 @@ 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_float64("falcon_h1.key_multiplier", self.hparams["key_multiplier"])
## Validation ##
assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported"
@ -6673,17 +6672,16 @@ class FalconH1Model(Mamba2Model):
# Add Falcon Mamba2 specific configuration
self.gguf_writer.add_uint32("falcon_h1.attention.head_dim", self.hparams["head_dim"])
self.gguf_writer.add_uint32("falcon_h1.ssm.mamba_d_inner", self.hparams["mamba_d_ssm"])
self.gguf_writer.add_ssm_head_dim(self.hparams["mamba_d_head"])
self.gguf_writer.add_ssm_inner_size(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"]))
self.gguf_writer.add_head_count(self.find_hparam(["num_attention_heads"]))
self.gguf_writer.add_key_length(self.hparams["head_dim"])
self.gguf_writer.add_value_length(self.hparams["head_dim"])
self.gguf_writer.add_head_count_kv(self.find_hparam(["num_key_value_heads"], optional=True) or
self.find_hparam(["num_attention_heads"]))
# Add any other Falcon Mamba2 specific configuration
self.gguf_writer.add_bool("falcon_h1.mamba_rms_norm", self.find_hparam(["mamba_rms_norm"], optional=True))
self.gguf_writer.add_rope_freq_base(self.find_hparam(["rope_theta"]))
###### CONVERSION LOGIC ######

3
src/llama-arch.cpp
View File

@ -128,8 +128,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" },
{ LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" },
{ LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" },
{ LLM_KV_ATTN_HEAD_DIM, "%s.attention.head_dim" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
{ LLM_KV_ATTENTION_MAX_ALIBI_BIAS, "%s.attention.max_alibi_bias" },

1
src/llama-model.cpp
View File

@ -1560,7 +1560,6 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state);
ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group);
ml.get_key(LLM_KV_SSM_HEAD_DIM, hparams.ssm_head_dim);
std::fill(hparams.recurrent_layer_arr.begin(), hparams.recurrent_layer_arr.end(), true);