llm : add MPT support (#3417)

* CUDA: added support for ggml_clamp (see also: ggml-org/ggml#545)

* mpt : added an implementation based (mostly) on falcon integration, modified with deltas from ggml/examples/mpt

* mpt : protect against "clip_qkv": null in mpt-7b

* mpt : quick fix to avoid "Strange model" warning when quantizing MPT models

* mpt : addendum to changeset:84e30e8 - leave parameter clamp_kqv out from metadata rather than use 0.0 to indicate "no clamping" (more compliant with the current GGUF spec?)

* mpt : standardized all tensor names to follow GGUF spec

* mpt : addendum to changeset:1be89c40 - use "req" parameter of GGUF_GET_KEY macro instead of duplicate code

* mpt : fixed comment s/gptneox/mpt/

* mpt : remove tabs, trailing whitespace

* mpt : removed ne01 + n_past == ne00 assertion from alibi (cuda/f32) and rope_shift from build_mpt

* mpt : updated convert-mpt-hf-to-gguf.py to reflect changes made to convert-gptneox-hf-to-gguf.py in pr:3252

* comment out n_past instead of marking it unused

* mpt : removed hardcoded +178 from convert script in favor of utilizing hparams["vocab_size"]

* mpt : remove unused tokenizer_json in convert script

* ggml : remove obsolete n_past assert in ggml_alibi

* llama : print clam_kqv and max_alibi_bias hparams

---------

Co-authored-by: Cebtenzzre <[email protected]>
Co-authored-by: Georgi Gerganov <[email protected]>

Author: 3 people

convert-mpt-hf-to-gguf.py

@@ -0,0 +1,216 @@
+#!/usr/bin/env python3
+# HF mpt--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert an MPT model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    parser.add_argument(
+        "--outfile", type=Path,
+        help="path to write to; default: based on input",
+    )
+    parser.add_argument(
+        "model", type=Path,
+        help="directory containing model file, or model file itself (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "MPTForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.MPT
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layers"]
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_context_length(hparams["max_seq_len"])
+gguf_writer.add_embedding_length(hparams["d_model"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
+gguf_writer.add_head_count(hparams["n_heads"])
+gguf_writer.add_layer_norm_eps(1e-05)
+if hparams["attn_config"]["clip_qkv"] is not None:
+    gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])
+gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"])
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# MPT token embedding tensors have dimension 50432 (hparams["vocab_size"]), but
+# there are only 50254 (len(tokenizer.vocab)) tokens in the vocab, presumably to
+# accomodate some "reserved" tokens; this is causing problems down the line in
+# llama.cpp, so we pad the vocab with dummy tokens:
+
+vocab_size = hparams["vocab_size"]
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i] if i in reverse_vocab else f"[PAD{i}]")
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True)
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Cannot map tensor '" + name + "'")
+            continue # for the sake of compatibility with some old published models, don't quit
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if ftype == 0 and data_dtype == np.float16:
+            data = data.astype(np.float32)
+
+        # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+        if ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+        # note: MPT output is tied to (same as) wte in original model;
+        # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
+        if new_name == "token_embd.weight":
+            gguf_writer.add_tensor("output.weight", data)
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

ggml-cuda.cu

@@ -415,6 +415,7 @@ static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_
 #define CUDA_SILU_BLOCK_SIZE 256
 #define CUDA_CPY_BLOCK_SIZE 32
 #define CUDA_SCALE_BLOCK_SIZE 256
+#define CUDA_CLAMP_BLOCK_SIZE 256
 #define CUDA_ROPE_BLOCK_SIZE 256
 #define CUDA_ALIBI_BLOCK_SIZE 32
 #define CUDA_DIAG_MASK_INF_BLOCK_SIZE 32
@@ -4585,6 +4586,15 @@ static __global__ void scale_f32(const float * x, float * dst, const float scale
     dst[i] = scale * x[i];
 }
 
+static __global__ void clamp_f32(const float * x, float * dst, const float min, const float max, const int k) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    dst[i] = x[i] < min ? min : (x[i] > max ? max : x[i]);
+}
 
 template<int qk, int qr, dequantize_kernel_t dq>
 static void get_rows_cuda(const void * x, const int32_t * y, float * dst, const int nrows, const int ncols, cudaStream_t stream) {
@@ -5475,6 +5485,11 @@ static void scale_f32_cuda(const float * x, float * dst, const float scale, cons
     scale_f32<<<num_blocks, CUDA_SCALE_BLOCK_SIZE, 0, stream>>>(x, dst, scale, k);
 }
 
+static void clamp_f32_cuda(const float * x, float * dst, const float min, const float max, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_CLAMP_BLOCK_SIZE - 1) / CUDA_CLAMP_BLOCK_SIZE;
+    clamp_f32<<<num_blocks, CUDA_CLAMP_BLOCK_SIZE, 0, stream>>>(x, dst, min, max, k);
+}
+
 template<typename T>
 static void rope_cuda(const T * x, T * dst, const int ncols, const int nrows, const int32_t * pos, const float freq_scale,
                           const int p_delta_rows, const float theta_scale, cudaStream_t stream) {
@@ -6419,12 +6434,12 @@ inline void ggml_cuda_op_alibi(
     const int64_t ne02 = src0->ne[2];
     const int64_t nrows = ggml_nrows(src0);
 
-    const int n_past = ((int32_t *) dst->op_params)[0];
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_head = ((int32_t *) dst->op_params)[1];
     float max_bias;
     memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-    GGML_ASSERT(ne01 + n_past == ne00);
+    //GGML_ASSERT(ne01 + n_past == ne00);
     GGML_ASSERT(n_head == ne02);
 
     const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@@ -6500,6 +6515,24 @@ inline void ggml_cuda_op_scale(
     (void) src1_dd;
 }
 
+inline void ggml_cuda_op_clamp(
+    const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
+    const float * src0_dd, const float * src1_dd, float * dst_dd, const cudaStream_t & main_stream) {
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    const float min = ((float *) dst->op_params)[0];
+    const float max = ((float *) dst->op_params)[1];
+
+    clamp_f32_cuda(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
+    CUDA_CHECK(cudaGetLastError());
+
+    (void) src1;
+    (void) dst;
+    (void) src1_dd;
+}
+
 static void ggml_cuda_op_flatten(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const ggml_cuda_op_flatten_t op) {
     const int64_t nrows0 = ggml_nrows(src0);
 
@@ -7061,6 +7094,10 @@ static void ggml_cuda_scale(const ggml_tensor * src0, const ggml_tensor * src1,
     ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_scale);
 }
 
+static void ggml_cuda_clamp(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    ggml_cuda_op_flatten(src0, src1, dst, ggml_cuda_op_clamp);
+}
+
 static void ggml_cuda_cpy(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne = ggml_nelements(src0);
     GGML_ASSERT(ne == ggml_nelements(src1));
@@ -7470,6 +7507,12 @@ bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_
         case GGML_OP_SCALE:
             func = ggml_cuda_scale;
             break;
+        case GGML_OP_CLAMP:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cuda_clamp;
+            break;
         case GGML_OP_CPY:
             func = ggml_cuda_cpy;
             break;

ggml-metal.m

@@ -1299,7 +1299,7 @@ void ggml_metal_graph_compute(
 
                             const int nth = MIN(1024, ne00);
 
-                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+                            //const int n_past = ((int32_t *) dst->op_params)[0];
                             const int n_head = ((int32_t *) dst->op_params)[1];
                             float max_bias;
                             memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));

ggml.c

@@ -13059,13 +13059,11 @@ static void ggml_compute_forward_alibi_f32(
         return;
     }
 
-    const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+    //const int n_past = ((int32_t *) dst->op_params)[0];
     const int n_head = ((int32_t *) dst->op_params)[1];
     float max_bias;
     memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
 
-    assert(n_past >= 0);
-
     const int64_t ne0 = src0->ne[0]; // all_seq_len = n_past + ne1
     const int64_t ne1 = src0->ne[1]; // seq_len_without_past
     const int64_t ne2 = src0->ne[2]; // n_head -> this is k