Refactor lora adapter support (ggerganov#8332)

* lora: load to devide buft

* add patch tensor function

* correct tensor patch

* llama_lora_adapter_apply

* correct ggml_backend_tensor_copy

* add llm_build_mm

* fix auto merge

* update based on review comments

* add convert script

* no more transpose A

* add f16 convert

* add metadata check

* add sanity check

* fix ftype

* add requirements

* fix requirements

* fix outfile

* conversion: only allow selected models

* fix types

* cuda : do not use dmmv if the tensor does not have enough cols

* llama : lora fixes

* do not disable mmap with lora

Co-authored-by: slaren <slarengh@gmail.com>

* llm_build_lora_mm_id

* convert_lora : MoE LoRA conversion support

* convert_lora : prefer safetensors, similarly to convert_hf

* convert_hf : simplify modify_tensors for InternLM2

* convert_lora : lazy conversion

* llama : load and use alpha from LoRA adapters

* llama : use llm_build_lora_mm in most model graphs

* auto scale

* Revert "auto scale"

This reverts commit 42415a4.

* remove redundant params

* Apply suggestions from code review

Co-authored-by: slaren <slarengh@gmail.com>

* change kv metadata

* move add_type to __init__

* convert_hf : move add_type to main()

* convert_lora : use the GGUFWriter from Model instead of overwriting it

---------

Co-authored-by: slaren <slarengh@gmail.com>
Co-authored-by: Francis Couture-Harpin <git@compilade.net>

common/common.cpp

@@ -685,15 +685,13 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa
     if (arg == "--lora") {
         CHECK_ARG
         params.lora_adapter.emplace_back(argv[i], 1.0f);
-        params.use_mmap = false;
         return true;
     }
     if (arg == "--lora-scaled") {
         CHECK_ARG
         const char* lora_adapter = argv[i];
         CHECK_ARG
         params.lora_adapter.emplace_back(lora_adapter, std::stof(argv[i]));
-        params.use_mmap = false;
         return true;
     }
     if (arg == "--lora-base") {
@@ -2089,19 +2087,14 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
         const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
         float lora_scale = std::get<1>(params.lora_adapter[i]);
-        int err = llama_model_apply_lora_from_file(model,
-                                             lora_adapter.c_str(),
-                                             lora_scale,
-                                             ((i > 0) || params.lora_base.empty())
-                                                ? NULL
-                                                : params.lora_base.c_str(),
-                                             params.n_threads);
-        if (err != 0) {
+        auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
+        if (adapter == nullptr) {
             fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
             llama_free(lctx);
             llama_free_model(model);
             return std::make_tuple(nullptr, nullptr);
         }
+        llama_lora_adapter_set(lctx, adapter, lora_scale);
     }
 
     if (params.ignore_eos) {

convert_hf_to_gguf.py

@@ -2264,13 +2264,6 @@ def set_vocab(self):
 
         special_vocab.add_to_gguf(self.gguf_writer)
 
-    def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int):
-        if n_head_kv is not None and n_head != n_head_kv:
-            n_head = n_head_kv
-        return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-                .swapaxes(1, 2)
-                .reshape(weights.shape))
-
     def set_gguf_parameters(self):
         self.gguf_writer.add_name("InternLM2")
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
@@ -2290,26 +2283,22 @@ def set_gguf_parameters(self):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         num_heads = self.hparams["num_attention_heads"]
         num_kv_heads = self.hparams["num_key_value_heads"]
-        hidden_size = self.hparams["hidden_size"]
+        n_embd = self.hparams["hidden_size"]
         q_per_kv = num_heads // num_kv_heads
-        head_dim = hidden_size // num_heads
+        head_dim = n_embd // num_heads
         num_groups = num_heads // q_per_kv
 
-        qkv_pattern = r"model\.layers\.(\d+)\.attention\.wqkv"
-
-        if re.match(qkv_pattern, name):
-            bid = re.findall(qkv_pattern, name)[0]
+        if bid is not None and f"model.layers.{bid}.attention.wqkv" in name:
             qkv = data_torch
-            # qkv = rearrange(qkv.T, " o (g n i) ->o g n i", g=num_groups, n=q_per_kv + 2, i=head_dim)
-            qkv = qkv.T.reshape((-1, num_groups, q_per_kv + 2, head_dim))
-            q, k, v = qkv[..., : q_per_kv, :], qkv[..., q_per_kv: q_per_kv + 1, :], qkv[..., q_per_kv + 1: q_per_kv + 2, :]
+
+            qkv = qkv.reshape((num_groups, q_per_kv + 2, head_dim, n_embd))
+            q, k, v = qkv[:, : q_per_kv], qkv[:, -2], qkv[:, -1]
+
             # The model weights of q and k equire additional reshape.
-            # q = self._hf_permute_qk(rearrange(q, " o g n i ->  o (g n i)").T, num_heads, num_heads)
-            q = self._hf_permute_qk(q.reshape((q.shape[0], -1)).T, num_heads, num_heads)
-            # k = self._hf_permute_qk(rearrange(k, " o g n i ->  o (g n i)").T, num_heads, num_kv_heads)
-            k = self._hf_permute_qk(k.reshape((k.shape[0], -1)).T, num_heads, num_kv_heads)
-            # v = rearrange(v, " o g n i ->  o (g n i)").T
-            v = v.reshape((v.shape[0], -1)).T
+            q = LlamaModel.permute(q.reshape((-1, q.shape[-1])), num_heads, num_heads)
+            k = LlamaModel.permute(k.reshape((-1, k.shape[-1])), num_heads, num_kv_heads)
+            v = v.reshape((-1, v.shape[-1]))
+
             return [
                 (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), q),
                 (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), k),
@@ -3585,6 +3574,7 @@ def main() -> None:
                                      small_first_shard=args.no_tensor_first_split)
 
         logger.info("Set model parameters")
+        model_instance.gguf_writer.add_type(gguf.GGUFType.MODEL)
         model_instance.set_gguf_parameters()
 
         logger.info("Set model tokenizer")