From dd15c0f93d3e416a4da0e4328024d343c7e1452f Mon Sep 17 00:00:00 2001
From: FFengIll <fengyouzheng@gmail.com>
Date: Thu, 14 Sep 2023 18:55:47 +0800
Subject: [PATCH 1/4] bugfix: `type_vocab_size` is also a hparam (can not use
const as 2).
---
bert.cpp | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/bert.cpp b/bert.cpp
index 2e033ba..d2dc0bd 100644
--- a/bert.cpp
+++ b/bert.cpp
@@ -23,6 +23,7 @@ struct bert_hparams
int32_t n_intermediate = 1536;
int32_t n_head = 12;
int32_t n_layer = 6;
+ int32_t n_vocab_size = 2;
int32_t f16 = 1;
};
@@ -489,11 +490,13 @@ struct bert_ctx * bert_load_from_file(const char *fname)
const int n_intermediate = hparams.n_intermediate;
const int n_max_tokens = hparams.n_max_tokens;
const int n_vocab = hparams.n_vocab;
+ const int n_vocab_size = hparams.n_vocab_size;
+
model.layers.resize(n_layer);
model.word_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
- model.token_type_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, 2);
+ model.token_type_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab_size);
model.position_embeddings = ggml_new_tensor_2d(ctx, wtype, n_embd, n_max_tokens);
model.ln_e_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
From a3c8548ec16213aea2159beb84bd355cfcd32b97 Mon Sep 17 00:00:00 2001
From: FFengIll <fengyouzheng@gmail.com>
Date: Thu, 14 Sep 2023 18:56:09 +0800
Subject: [PATCH 2/4] bugfix: add `type_vocab_size`.
---
models/convert-to-ggml.py | 1 +
1 file changed, 1 insertion(+)
diff --git a/models/convert-to-ggml.py b/models/convert-to-ggml.py
index 7ef5b80..c05ee25 100644
--- a/models/convert-to-ggml.py
+++ b/models/convert-to-ggml.py
@@ -61,6 +61,7 @@
fout.write(struct.pack("i", hparams["intermediate_size"]))
fout.write(struct.pack("i", hparams["num_attention_heads"]))
fout.write(struct.pack("i", hparams["num_hidden_layers"]))
+fout.write(struct.pack("i", hparams["type_vocab_size"]))
fout.write(struct.pack("i", ftype))
for i in range(hparams["vocab_size"]):
From 7ef3126bb8f58a26f54ed36743c0f1d1e16954f9 Mon Sep 17 00:00:00 2001
From: FFengIll <fengyouzheng@gmail.com>
Date: Tue, 19 Sep 2023 10:35:43 +0800
Subject: [PATCH 3/4] bugfix: missing read for new n_vocab_size.
---
bert.cpp | 2 ++
1 file changed, 2 insertions(+)
diff --git a/bert.cpp b/bert.cpp
index d2dc0bd..f04c6c5 100644
--- a/bert.cpp
+++ b/bert.cpp
@@ -365,6 +365,7 @@ struct bert_ctx * bert_load_from_file(const char *fname)
fin.read((char *)&hparams.n_intermediate, sizeof(hparams.n_intermediate));
fin.read((char *)&hparams.n_head, sizeof(hparams.n_head));
fin.read((char *)&hparams.n_layer, sizeof(hparams.n_layer));
+ fin.read((char *)&hparams.n_vocab_size, sizeof(hparams.n_vocab_size));
fin.read((char *)&hparams.f16, sizeof(hparams.f16));
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
@@ -373,6 +374,7 @@ struct bert_ctx * bert_load_from_file(const char *fname)
printf("%s: n_intermediate = %d\n", __func__, hparams.n_intermediate);
printf("%s: n_head = %d\n", __func__, hparams.n_head);
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+ printf("%s: n_vocab_size = %d\n", __func__, hparams.n_vocab_size);
printf("%s: f16 = %d\n", __func__, hparams.f16);
}
From 79825d8ea8c387a28ecf778d26e61563a6eda57e Mon Sep 17 00:00:00 2001
From: F <fengyouzheng@gmail.com>
Date: Thu, 21 Sep 2023 04:25:39 +0000
Subject: [PATCH 4/4] bugfix: add n_vocab_size into quantize.
---
models/quantize.cpp | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/models/quantize.cpp b/models/quantize.cpp
index 22411a1..83f2fd2 100644
--- a/models/quantize.cpp
+++ b/models/quantize.cpp
@@ -20,6 +20,7 @@ struct bert_hparams
int32_t n_intermediate = 1536;
int32_t n_head = 12;
int32_t n_layer = 6;
+ int32_t n_vocab_size = 2;
int32_t f16 = 1;
};
@@ -74,6 +75,7 @@ bool bert_model_quantize(const std::string & fname_inp, const std::string & fnam
finp.read((char *) &hparams.n_intermediate, sizeof(hparams.n_intermediate));
finp.read((char *) &hparams.n_head, sizeof(hparams.n_head));
finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+ finp.read((char *) &hparams.n_vocab_size, sizeof(hparams.n_vocab_size));
finp.read((char *) &hparams.f16, sizeof(hparams.f16));
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
@@ -82,6 +84,7 @@ bool bert_model_quantize(const std::string & fname_inp, const std::string & fnam
printf("%s: n_intermediate = %d\n", __func__, hparams.n_intermediate);
printf("%s: n_head = %d\n", __func__, hparams.n_head);
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+ printf("%s: n_vocab_size = %d\n", __func__, hparams.n_vocab_size);
printf("%s: f16 = %d\n", __func__, hparams.f16);
fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
@@ -90,6 +93,7 @@ bool bert_model_quantize(const std::string & fname_inp, const std::string & fnam
fout.write((char *) &hparams.n_intermediate, sizeof(hparams.n_intermediate));
fout.write((char *) &hparams.n_head, sizeof(hparams.n_head));
fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+ fout.write((char *) &hparams.n_vocab_size, sizeof(hparams.n_vocab_size));
fout.write((char *) &itype, sizeof(hparams.f16));
}