fixed multipliers

Author: ibrahimkhadraoui

convert_hf_to_gguf.py

@@ -6645,11 +6645,11 @@ def set_gguf_parameters(self):
         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_float64("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_float64("falcon_h1.lm_head_multiplier", self.hparams["lm_head_multiplier"])
+        self.gguf_writer.add_float64("falcon_h1.embedding_multiplier", self.hparams["embedding_multiplier"])
 
         ## Validation ##
         assert self.hparams.get("hidden_act") in [None, "silu"], "Only SILU activation supported"
@@ -6666,15 +6666,15 @@ def set_gguf_parameters(self):
                                     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_float64("falcon_h1.attention_in_multiplier", self.attention_in_multiplier)
+        self.gguf_writer.add_float64("falcon_h1.attention_out_multiplier", self.attention_out_multiplier)
+        self.gguf_writer.add_float64("falcon_h1.ssm_in_multiplier", self.ssm_in_multiplier)
+        self.gguf_writer.add_float64("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_float64("falcon_h1.mlp_gate_multiplier", self.mlp_multipliers[0])
+            self.gguf_writer.add_float64("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:
@@ -6684,7 +6684,7 @@ def set_gguf_parameters(self):
         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_float64("falcon_h1.rope_theta", self.find_hparam(["rope_theta"], optional=True))
 
 ###### CONVERSION LOGIC ######
 

src/llama-hparams.h

@@ -122,17 +122,17 @@ struct llama_hparams {
     bool     mamba_use_mlp            = false;
     bool     mamba_norm_before_gate   = false;
     bool     mamba_rms_norm           = false;
-    float    attention_in_multiplier  = 1.0f;
-    float    attention_out_multiplier = 1.0f;
-    float    ssm_in_multiplier        = 1.0f;
-    float    ssm_out_multiplier       = 1.0f;
-    float    mlp_gate_multiplier      = 1.0f;
-    float    mlp_down_multiplier      = 1.0f;
-    float    key_multiplier           = 1.0f;
-    float    lm_head_multiplier       = 1.0f;
-    float    rope_theta               = 10000.0f;
+    double attention_in_multiplier  = 1.0;
+    double attention_out_multiplier = 1.0;
+    double ssm_in_multiplier        = 1.0;
+    double ssm_out_multiplier       = 1.0;
+    double mlp_gate_multiplier      = 1.0;
+    double mlp_down_multiplier      = 1.0;
+    double key_multiplier           = 1.0;
+    double lm_head_multiplier       = 1.0;
+    double rope_theta               = 10000.0;
+    double embedding_multiplier     = 1.0;
     bool     ssm_has_mup              = false;
-    float    embedding_multiplier     = 1.0f;
     uint32_t vocab_size               = 0;
     uint32_t intermediate_size        = 0;
     float    mamba_expand             = 0.0f;

src/llama-model-loader.cpp

@@ -400,6 +400,7 @@ namespace GGUFMeta {
 
     template bool llama_model_loader::get_key<bool>       (enum llm_kv kid, bool & result,        bool required);
     template bool llama_model_loader::get_key<float>      (enum llm_kv kid, float & result,       bool required);
+    template bool llama_model_loader::get_key<double>      (enum llm_kv kid, double & result,       bool required);
     template bool llama_model_loader::get_key<uint32_t>   (enum llm_kv kid, uint32_t & result,    bool required);
     template bool llama_model_loader::get_key<std::string>(enum llm_kv kid, std::string & result, bool required);