convert_hf : simplify modify_tensors for InternLM2

* convert_lora : lazy conversion

* llama : load and use alpha from LoRA adapters

convert_hf_to_gguf.py

@@ -2222,13 +2222,6 @@ class InternLM2Model(Model):
 
         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"])
@@ -2248,26 +2241,22 @@ class InternLM2Model(Model):
     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),

convert_lora_to_gguf.py

@@ -8,9 +8,10 @@ import logging
 import argparse
 import os
 import sys
+import json
+from math import prod
 from pathlib import Path
-from types import EllipsisType
-from typing import TYPE_CHECKING, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
+from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
 
 import torch
 
@@ -22,7 +23,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 # reuse model definitions from convert_hf_to_gguf.py
-from convert_hf_to_gguf import Model
+from convert_hf_to_gguf import LazyTorchTensor, Model
 
 logger = logging.getLogger("lora-to-gguf")
 
@@ -35,37 +36,45 @@ class PartialLoraTensor:
 
 # magic to support tensor shape modifications and splitting
 class LoraTorchTensor:
-    _lora_A: Tensor
-    _lora_B: Tensor
+    _lora_A: Tensor  # (n_rank, row_size)
+    _lora_B: Tensor  # (col_size, n_rank)
     _rank: int
 
     def __init__(self, A: Tensor, B: Tensor):
         assert len(A.shape) == len(B.shape)
+        assert A.shape[-2] == B.shape[-1]
         if A.dtype != B.dtype:
             A = A.to(torch.float32)
             B = B.to(torch.float32)
         self._lora_A = A
         self._lora_B = B
-        assert self._lora_A.shape[-2] == self._lora_B.shape[-1]
-        self._rank = self._lora_B.shape[-1]
+        self._rank = B.shape[-1]
+
+    def get_lora_A_B(self) -> tuple[Tensor, Tensor]:
+        return (self._lora_A, self._lora_B)
 
     def __getitem__(
         self,
         indices: (
             SupportsIndex
             | slice
-            | tuple[SupportsIndex | slice | EllipsisType | Tensor, ...]
+            | tuple[SupportsIndex | slice | Tensor, ...]  # TODO: add ellipsis in the type signature
         ),
     ) -> LoraTorchTensor:
         shape = self.shape
-        if isinstance(indices, (SupportsIndex, slice)):
+        if isinstance(indices, SupportsIndex):
             if len(shape) > 2:
                 return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
             else:
-                raise NotImplementedError
+                raise NotImplementedError  # can't return a vector
+        elif isinstance(indices, slice):
+            if len(shape) > 2:
+                return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
+            else:
+                return LoraTorchTensor(self._lora_A, self._lora_B[indices])
         elif isinstance(indices, tuple):
             assert len(indices) > 0
-            if isinstance(indices[-1], EllipsisType):
+            if indices[-1] is Ellipsis:
                 return self[indices[:-1]]
             # expand ellipsis
             indices = tuple(
@@ -73,7 +82,7 @@ class LoraTorchTensor:
                 for v in (
                     (
                         (slice(None, None) for _ in range(len(indices) - 1))
-                        if isinstance(i, EllipsisType)
+                        if i is Ellipsis
                         else (i,)
                     )
                     for i in indices
@@ -85,11 +94,14 @@ class LoraTorchTensor:
                 indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape))))
 
             # TODO: make sure this is correct
-            # lora_A has a shape which looks like (..., 1, 1, rank, self.shape[-1])
             indices_A = (
                 *(
-                    0 if isinstance(i, SupportsIndex) else slice(None, None)
-                    for i in indices[:-2]
+                    (
+                        j.__index__() % self._lora_A.shape[i]
+                        if isinstance(j, SupportsIndex)
+                        else slice(None, None)
+                    )
+                    for i, j in enumerate(indices[:-2])
                 ),
                 slice(None, None),
                 indices[-1],
@@ -97,7 +109,7 @@ class LoraTorchTensor:
             indices_B = indices[:-1]
             return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B])
         else:
-            raise NotImplementedError
+            raise NotImplementedError  # unknown indice type
 
     @property
     def dtype(self) -> torch.dtype:
@@ -106,23 +118,37 @@ class LoraTorchTensor:
 
     @property
     def shape(self) -> tuple[int, ...]:
+        assert len(self._lora_A.shape) == len(self._lora_B.shape)
         return (*self._lora_B.shape[:-1], self._lora_A.shape[-1])
 
     def size(self, dim=None):
         assert dim is None
         return self.shape
 
-    def reshape(self, *shape: int | tuple[int]) -> LoraTorchTensor:
+    def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
         if isinstance(shape[0], tuple):
-            new_shape: tuple[int] = shape[0]
+            new_shape: tuple[int, ...] = shape[0]
         else:
-            new_shape = cast(tuple[int], shape)
+            new_shape = cast(tuple[int, ...], shape)
         orig_shape = self.shape
+        if len(new_shape) < 2:
+            raise NotImplementedError  # can't become a vector
+
+        # expand -1 in the shape
+        if any(dim == -1 for dim in new_shape):
+            n_elems = prod(orig_shape)
+            n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape)
+            assert n_elems % n_new_elems == 0
+            new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),)
+
         if new_shape[-1] != orig_shape[-1]:
-            raise NotImplementedError
+            raise NotImplementedError  # can't reshape the row size trivially
+
+        shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1])
+        shape_B = (*new_shape[:-1], self._rank)
         return LoraTorchTensor(
-            self._lora_A.reshape((*(1 for _ in new_shape[:-2]), *self._lora_A.shape[-2:])),
-            self._lora_B.reshape((*new_shape[:-1], self._rank)),
+            self._lora_A.reshape(shape_A),
+            self._lora_B.reshape(shape_B),
         )
 
     def reshape_as(self, other: Tensor) -> LoraTorchTensor:
@@ -134,12 +160,15 @@ class LoraTorchTensor:
     def permute(self, *dims: int) -> LoraTorchTensor:
         shape = self.shape
         dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims)
-        if dims[-1] == -2 and dims[-2] == -1:
-            return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims))
-        else:
-            assert dims[-1] == -1
+        if dims[-1] == -1:
+            # TODO: support higher dimensional A shapes bigger than 1
             assert all(dim == 1 for dim in self._lora_A.shape[:-2])
             return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims))
+        if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1:
+            return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims))
+        else:
+            # TODO: compose the above two
+            raise NotImplementedError
 
     def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor:
         shape = self.shape
@@ -181,11 +210,13 @@ class LoraTorchTensor:
                     torch.cat([a._lora_A for a in args[0]], dim),
                     torch.cat([b._lora_B for b in args[0]], dim),
                 )
-            else:
+            elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]):
                 return LoraTorchTensor(
-                    args[0][0]._lora_A,  # TODO: is this correct? (can't cat over the rank)
+                    args[0][0]._lora_A,
                     torch.cat([b._lora_B for b in args[0]], dim),
                 )
+            else:
+                raise NotImplementedError
         else:
             raise NotImplementedError
 
@@ -205,13 +236,17 @@ def parse_args() -> argparse.Namespace:
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
     )
     parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0"], default="f16",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",
         help="model is executed on big endian machine",
     )
+    parser.add_argument(
+        "--no-lazy", action="store_true",
+        help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)",
+    )
     parser.add_argument(
         "--verbose", action="store_true",
         help="increase output verbosity",
@@ -237,13 +272,16 @@ if __name__ == '__main__':
         "f16": gguf.LlamaFileType.MOSTLY_F16,
         "bf16": gguf.LlamaFileType.MOSTLY_BF16,
         "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+        "auto": gguf.LlamaFileType.GUESSED,
     }
+
     ftype = ftype_map[args.outtype]
 
-    dir_base_model = args.base
-    dir_lora = args.lora_path
-    input_json = os.path.join(dir_lora, "adapter_config.json")
-    input_model = os.path.join(dir_lora, "adapter_model.safetensors")
+    dir_base_model: Path = args.base
+    dir_lora: Path = args.lora_path
+    lora_config = dir_lora / "adapter_config.json"
+    input_model = dir_lora / "adapter_model.safetensors"
+
     if args.outfile is not None:
         fname_out = args.outfile
     else:
@@ -276,6 +314,8 @@ if __name__ == '__main__':
                 tensor_map: dict[str, PartialLoraTensor] = {}
 
                 for name, tensor in lora_model.items():
+                    if self.lazy:
+                        tensor = LazyTorchTensor.from_eager(tensor)
                     base_name = get_base_tensor_name(name)
                     is_lora_a = ".lora_A.weight" in name
                     is_lora_b = ".lora_B.weight" in name
@@ -305,16 +345,30 @@ if __name__ == '__main__':
                 dest = super().modify_tensors(data_torch, name, bid)
                 for dest_name, dest_data in dest:
                     assert isinstance(dest_data, LoraTorchTensor)
-                    # logger.info(f"{orig_name} --> {dest_name}")
-                    yield (dest_name + ".lora_a", dest_data._lora_A)
-                    yield (dest_name + ".lora_b", dest_data._lora_B)
+                    lora_a, lora_b = dest_data.get_lora_A_B()
 
-        model_instance = LoraModel(dir_base_model, ftype, fname_out, args.bigendian, False, False, None)
+                    yield (dest_name + ".lora_a", lora_a)
+                    yield (dest_name + ".lora_b", lora_b)
+
+        model_instance = LoraModel(
+            dir_base_model,
+            ftype,
+            fname_out,
+            is_big_endian=args.bigendian,
+            use_temp_file=False,
+            eager=args.no_lazy,
+            model_name=None,
+        )
         logger.info("Set model parameters")
         model_instance.set_gguf_parameters()
 
-        # adapter_config = json.load(input_json)
+        with open(lora_config, "r") as f:
+            lparams: dict[str, Any] = json.load(f)
+
+        alpha = lparams["lora_alpha"]
+
         model_instance.gguf_writer.add_string("training.type", "finetune_lora")
+        model_instance.gguf_writer.add_float32("training.lora.alpha", float(alpha))
 
         model_instance.gguf_writer.add_quantization_version(gguf.GGML_QUANT_VERSION)
         logger.info("Exporting model...")

gguf-py/gguf/quants.py

@@ -43,7 +43,7 @@ def __apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.
         osize *= dim
     out = np.empty(shape=osize, dtype=otype)
     # compute over groups of 16 rows (arbitrary, but seems good for performance)
-    n_groups = rows.shape[0] // 16
+    n_groups = (rows.shape[0] // 16) or 1
     np.concatenate([func(group).ravel() for group in np.array_split(rows, n_groups)], axis=0, out=out)
     return out.reshape(oshape)
 

src/llama.cpp

@@ -379,6 +379,7 @@ enum llm_kv {
     LLM_KV_TOKENIZER_EOT_ID,
 
     LLM_KV_TRAINING_TYPE,
+    LLM_KV_TRAINING_LORA_ALPHA,
 };
 
 static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
@@ -473,7 +474,8 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_TOKENIZER_MIDDLE_ID,            "tokenizer.ggml.middle_token_id"          },
     { LLM_KV_TOKENIZER_EOT_ID,               "tokenizer.ggml.eot_token_id"             },
 
-    { LLM_KV_TRAINING_TYPE,                  "training.type" },
+    { LLM_KV_TRAINING_TYPE,                  "training.type"       },
+    { LLM_KV_TRAINING_LORA_ALPHA,            "training.lora.alpha" },
 };
 
 struct LLM_KV {
@@ -2848,6 +2850,8 @@ struct llama_lora_adapter {
     std::vector<struct ggml_context *> ctxs;
     std::vector<ggml_backend_buffer_t> bufs;
 
+    float alpha;
+
     llama_lora_adapter(struct llama_model * base_model): base_model(base_model) {
         base_model->lora_adapters.insert(this);
     }
@@ -7878,10 +7882,12 @@ static struct ggml_tensor * llm_build_lora_mm(
     struct ggml_tensor * res = ggml_mul_mat(ctx0, w, cur);
     for (auto & it : lctx.lora_adapters) {
         struct llama_lora_weight * lora = it.first->get_weight(w);
-        float scale = it.second;
         if (lora == nullptr) {
             continue;
         }
+        const float alpha = it.first->alpha;
+        const float rank  = (float) lora->b->ne[0];
+        const float scale = alpha ? it.second * alpha / rank : it.second;
         struct ggml_tensor * ab_cur = ggml_mul_mat(
             ctx0, lora->b,
             ggml_mul_mat(ctx0, lora->a, cur)
@@ -7902,10 +7908,12 @@ static struct ggml_tensor * llm_build_lora_mm_id(
     struct ggml_tensor * res = ggml_mul_mat_id(ctx0, w, cur, ids);
     for (auto & it : lctx.lora_adapters) {
         struct llama_lora_weight * lora = it.first->get_weight(w);
-        float scale = it.second;
         if (lora == nullptr) {
             continue;
         }
+        const float alpha = it.first->alpha;
+        const float rank  = (float) lora->b->ne[0];
+        const float scale = alpha ? it.second * alpha / rank : it.second;
         struct ggml_tensor * ab_cur = ggml_mul_mat_id(
             ctx0, lora->b,
             ggml_mul_mat_id(ctx0, lora->a, cur, ids),
@@ -18587,10 +18595,14 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c
 
     // check metadata
     {
-        auto get_kv_str = [&](std::string key) -> std::string {
+        auto get_kv_str = [&](const std::string & key) -> std::string {
             int id = gguf_find_key(ctx_gguf, key.c_str());
             return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
         };
+        auto get_kv_f32 = [&](const std::string & key) -> float {
+            int id = gguf_find_key(ctx_gguf, key.c_str());
+            return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id);
+        };
         LLM_KV llm_kv = LLM_KV(LLM_ARCH_UNKNOWN);
         auto lora_arch_name = get_kv_str(llm_kv(LLM_KV_GENERAL_ARCHITECTURE));
         auto lora_arch = llm_arch_from_string(lora_arch_name);
@@ -18604,6 +18616,8 @@ static void llama_lora_adapter_init_internal(struct llama_model * model, const c
             gguf_free(ctx_gguf);
             throw std::runtime_error("expect training.type to be finetune_lora, but got: " + train_type);
         }
+
+        adapter.alpha = get_kv_f32(llm_kv(LLM_KV_TRAINING_LORA_ALPHA));
     }
 
     int n_tensors = gguf_get_n_tensors(ctx_gguf);