more cleanups and fixed conversion

Author: younesbelkada

convert_hf_to_gguf.py

@@ -6655,7 +6655,7 @@ def set_gguf_parameters(self):
         ## Mamba mixer params ##
         self.gguf_writer.add_ssm_conv_kernel(self.find_hparam(["conv_kernel", "d_conv"]))
         self.gguf_writer.add_ssm_group_count(self.n_group)
-        self.gguf_writer.add_ssm_inner_size(self.d_inner)
+        self.gguf_writer.add_ssm_inner_size(self.find_hparam(["mamba_d_ssm"]))
         self.gguf_writer.add_ssm_head_dim(d_head := self.find_hparam(["d_head"]))
         self.gguf_writer.add_ssm_time_step_rank(self.find_hparam(["n_heads"]))
 

src/llama-arch.cpp

@@ -1736,7 +1736,6 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
     {LLM_TENSOR_CLS,                        {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_CLS_OUT,                    {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}},
     {LLM_TENSOR_OUTPUT_NORM,                {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
-    {LLM_TENSOR_FINAL_NORM,                 {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
     {LLM_TENSOR_DEC_OUTPUT_NORM,            {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
     {LLM_TENSOR_ENC_OUTPUT_NORM,            {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL}},
     {LLM_TENSOR_ROPE_FREQS,                 {LLM_TENSOR_LAYER_REPEATING, GGML_OP_ROPE}},

src/llama-arch.h

@@ -390,7 +390,6 @@ enum llm_tensor {
     LLM_TENSOR_POS_NET_ATTN_K,
     LLM_TENSOR_POS_NET_ATTN_V,
     LLM_TENSOR_POS_NET_ATTN_OUT,
-    LLM_TENSOR_FINAL_NORM,
 };
 
 enum llm_tensor_layer {