more cleanups

Author: younesbelkada

convert_hf_to_gguf.py

@@ -6673,17 +6673,16 @@ def set_gguf_parameters(self):
 
 
         # 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 ######

src/llama-arch.cpp

@@ -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"               },

src/llama-arch.h

@@ -158,7 +158,6 @@ enum llm_kv {
     LLM_KV_ATTENTION_LAYER_INDICES,
 
     // Falcon-H1 specific
-    LLM_KV_ATTN_HEAD_DIM,
     LLM_KV_SSM_HEAD_DIM,
     LLM_KV_N_LAYER,
     LLM_KV_FALCON_H1_MAMBA_RMS_NORM,

src/llama-model.cpp

@@ -4550,7 +4550,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
                         layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, ssm_num_heads}, 0);
                         layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, ssm_num_heads}, 0);
                         // ssm_norm
-                        layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {ssm_intermediate_size / ssm_n_groups, ssm_n_groups}, 0);
+                        layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {ssm_intermediate_size / ssm_n_groups, ssm_n_groups}, 1);
                         // out_proj
                         layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {ssm_intermediate_size, hidden_size}, 0);