tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-18 00:27:31 +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_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
self.gguf_writer.add_key_length(self.hparams["head_dim"]) self.gguf_writer.add_key_length(self.hparams["head_dim"])
self.gguf_writer.add_value_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 ## Other params
self.gguf_writer.add_float32("falcon-h1.lm_head_multiplier", self.hparams["lm_head_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"]) self.gguf_writer.add_float32("falcon_h1.embedding_multiplier", self.hparams["embedding_multiplier"])
## Validation ## ## Validation ##
assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported" 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 # 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.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.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.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_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.num_key_value_heads",
self.find_hparam(["num_key_value_heads"], optional=True) or self.find_hparam(["num_key_value_heads"], optional=True) or
self.find_hparam(["num_attention_heads"])) self.find_hparam(["num_attention_heads"]))
# Add multipliers as metadata instead of tensors # 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_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.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_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.ssm_out_multiplier", self.ssm_out_multiplier)
# Add MLP multipliers # Add MLP multipliers
if isinstance(self.mlp_multipliers, (list, tuple)) and len(self.mlp_multipliers) == 2: 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_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_down_multiplier", self.mlp_multipliers[1])
# Add has MuP flag if SSM multipliers are present # Add has MuP flag if SSM multipliers are present
if self.ssm_multipliers is not None: 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 # 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_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_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_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_float32("falcon_h1.rope_theta", self.find_hparam(["rope_theta"], optional=True))
###### CONVERSION LOGIC ###### ###### CONVERSION LOGIC ######

2
gguf-py/gguf/tensor_mapping.py
View File

@ -1177,7 +1177,7 @@ class TensorNameMap:
), ),
MODEL_TENSOR.SSM_MUP_VEC: ( MODEL_TENSOR.SSM_MUP_VEC: (
"model.layers.{bid}.mamba.mup_vector", # falcon-h1 "model.layers.{bid}.mamba.mup_vector", # falcon_h1
), ),
MODEL_TENSOR.SSM_NORM: ( MODEL_TENSOR.SSM_NORM: (

2
src/llama-arch.cpp
View File

@ -46,7 +46,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
{ LLM_ARCH_STARCODER2, "starcoder2" }, { LLM_ARCH_STARCODER2, "starcoder2" },
{ LLM_ARCH_MAMBA, "mamba" }, { LLM_ARCH_MAMBA, "mamba" },
{ LLM_ARCH_MAMBA2, "mamba2" }, { LLM_ARCH_MAMBA2, "mamba2" },
{ LLM_ARCH_FALCON_H1, "falcon-h1" }, { LLM_ARCH_FALCON_H1, "falcon_h1" },
{ LLM_ARCH_XVERSE, "xverse" }, { LLM_ARCH_XVERSE, "xverse" },
{ LLM_ARCH_COMMAND_R, "command-r" }, { LLM_ARCH_COMMAND_R, "command-r" },
{ LLM_ARCH_COHERE2, "cohere2" }, { LLM_ARCH_COHERE2, "cohere2" },

2
src/llama-model.h
View File

@ -226,7 +226,7 @@ struct llama_layer {
struct ggml_tensor * ffn_down_enc = nullptr; struct ggml_tensor * ffn_down_enc = nullptr;
struct ggml_tensor * ffn_up_enc = nullptr; struct ggml_tensor * ffn_up_enc = nullptr;
// falcon-h1 // falcon_h1
struct ggml_tensor * ssm_in_b = nullptr; struct ggml_tensor * ssm_in_b = nullptr;
struct ggml_tensor * ssm_mup_vec = nullptr; struct ggml_tensor * ssm_mup_vec = nullptr;

5
src/llama-quant.cpp
View File

@ -315,6 +315,11 @@ static ggml_type llama_tensor_get_type(quantize_state_impl & qs, ggml_type new_t
else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) { else if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XXS) {
new_type = GGML_TYPE_IQ2_S; new_type = GGML_TYPE_IQ2_S;
} }
} else if (name.find("ssm_in.weight") != std::string::npos) {
// For mamba-based models it's better to not quantize the ssm-proj layers
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) {
new_type = GGML_TYPE_BF16;
}
} else if (name.find("attn_q.weight") != std::string::npos) { } else if (name.find("attn_q.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) { if (ftype == LLAMA_FTYPE_MOSTLY_IQ3_XS) {
new_type = GGML_TYPE_IQ3_XXS; new_type = GGML_TYPE_IQ3_XXS;

2
src/llama-vocab.cpp
View File

@ -1522,7 +1522,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
tokenizer_pre == "llama-v3" || tokenizer_pre == "llama-v3" ||
tokenizer_pre == "llama-bpe"|| tokenizer_pre == "llama-bpe"||
tokenizer_pre == "falcon3" || tokenizer_pre == "falcon3" ||
tokenizer_pre == "falcon-h1" || tokenizer_pre == "falcon_h1" ||
tokenizer_pre == "pixtral") { tokenizer_pre == "pixtral") {
pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3; pre_type = LLAMA_VOCAB_PRE_TYPE_LLAMA3;
ignore_merges = true; ignore_merges = true;