add chatglm3-6b model support huggingface model:

https://hf-mirror.com/THUDM/chatglm3-6b

Signed-off-by: XingXing Qiao <qiaoxx@dingdao.com>

convert-hf-to-gguf.py

@@ -79,7 +79,7 @@ class Model:
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, ".bin")
         self.hparams = Model.load_hparams(self.dir_model)
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
+        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
         if self.ftype == gguf.LlamaFileType.GUESSED:
@@ -2710,6 +2710,167 @@ class DeepseekV2Model(Model):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("ChatGLMModel")
+class ChatGLMModel(Model):
+    model_arch = gguf.MODEL_ARCH.CHATGLM
+
+    def set_vocab(self):
+        dir_model = self.dir_model
+        hparams = self.hparams
+        tokens: list[bytearray] = []
+        toktypes: list[int] = []
+        scores: list[float] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
+        vocab_size = hparams.get("padded_vocab_size", len(tokenizer.get_vocab()))
+        assert max(tokenizer.get_vocab().values()) < vocab_size
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.get_vocab().items()}
+
+        for token_id in range(vocab_size):
+            piece = tokenizer._convert_id_to_token(token_id)
+            if token_id == 0:
+                piece = "<unk>"
+            elif token_id == 1:
+                piece = "<bos>"
+            elif token_id == 2:
+                piece = "<eos>"
+
+            text = piece.encode("utf-8")
+            score = 0.0
+            if len(piece) != 0 and token_id < 64789:
+                score = tokenizer.tokenizer.sp_model.get_score(token_id)
+
+            if len(piece) == 0:
+                text = f"[PAD{token_id}]".encode("utf-8")
+
+            if token_id >= 64789:
+                toktype = SentencePieceTokenTypes.UNKNOWN
+                tokens.append(text)
+                scores.append(score)
+                toktypes.append(toktype)
+                continue
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.tokenizer.sp_model.is_unknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.tokenizer.sp_model.is_control(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.tokenizer.sp_model.is_unused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.tokenizer.sp_model.is_byte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens.append(text)
+            scores.append(score)
+            toktypes.append(toktype)
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_name("ChatGLM-6b-chat")
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_head_kv = self.hparams.get("multi_query_group_num", n_head)
+        self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
+        self.gguf_writer.add_embedding_length(n_embed)
+        self.gguf_writer.add_feed_forward_length(self.hparams.get("ffn_hidden_size", 4 * n_embed))
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(n_head)
+        self.gguf_writer.add_head_count_kv(n_head_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layernorm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_rope_dimension_count(64)
+        self.gguf_writer.add_add_bos_token(False)
+
+    def write_tensors(self):
+        block_count = self.hparams["num_layers"]
+        tensors = dict(self.get_tensors())
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        has_lm_head = True
+        n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
+        n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
+
+        for name, data_torch in tensors.items():
+            if name.endswith(".rotary_pos_emb.inv_freq"):
+                continue
+
+            if "lm_head.weight" not in tensors.keys() and "output.weight" not in tensors.keys():
+                has_lm_head = False
+
+            name = re.sub(r'transformer\.', '', name)
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
+                # Map bloom-style qkv_linear to gpt-style qkv_linear
+                # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+                # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+                qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
+                data = np.concatenate(
+                    (
+                        qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                        qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                        qkv_weights[:, 2, :, :].reshape((-1, n_embed)),
+                    ),
+                    axis=0,
+                )
+                print("re-format attention.linear_qkv.weight")
+            elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
+                qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
+                data = np.concatenate(
+                    (
+                        qkv_bias[:, 0, :].reshape((n_embed,)),
+                        qkv_bias[:, 1, :].reshape((n_embed,)),
+                        qkv_bias[:, 2, :].reshape((n_embed,)),
+                    ),
+                    axis=0,
+                )
+                print("re-format attention.linear_qkv.bias")
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.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 self.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 self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"=> {new_name}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+            if not has_lm_head and name == "word_embeddings.weight":
+                self.gguf_writer.add_tensor("output.weight", data)
+                print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
+
+
 ###### CONVERSION LOGIC ######
 
 

gguf-py/gguf/constants.py

@@ -148,6 +148,7 @@ class MODEL_ARCH(IntEnum):
     OLMO       = auto()
     ARCTIC     = auto()
     DEEPSEEK2  = auto()
+    CHATGLM    = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -236,6 +237,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.OLMO:           "olmo",
     MODEL_ARCH.ARCTIC:         "arctic",
     MODEL_ARCH.DEEPSEEK2:      "deepseek2",
+    MODEL_ARCH.CHATGLM:        "chatglm",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -805,6 +807,18 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.CHATGLM : [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.ROPE_FREQS,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 
@@ -842,6 +856,9 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.ROPE_FREQS,
         MODEL_TENSOR.ATTN_ROT_EMBD,
     ],
+    MODEL_ARCH.CHATGLM: [
+        MODEL_TENSOR.ROPE_FREQS,
+    ],
 }
 
 #

gguf-py/gguf/tensor_mapping.py

@@ -24,6 +24,7 @@ class TensorNameMap:
             "backbone.embedding",                        # mamba
             "backbone.embeddings",                       # mamba-hf
             "transformer.in_out_embed",                  # Grok
+            "embedding.word_embeddings",                 # chatglm
         ),
 
         # Token type embeddings
@@ -52,6 +53,7 @@ class TensorNameMap:
             "output",                    # llama-pth bloom internlm2
             "word_embeddings_for_head",  # persimmon
             "lm_head.linear",            # phi2
+            "output_layer",              # chatglm
         ),
 
         # Output norm
@@ -68,11 +70,13 @@ class TensorNameMap:
             "model.norm_f",                            # mamba-qbert
             "backbone.norm_f",                         # mamba
             "transformer.rms_norm",                    # Grok
+            "encoder.final_layernorm",                 # chatglm
         ),
 
         # Rope frequencies
         MODEL_TENSOR.ROPE_FREQS: (
             "rope.freqs",  # llama-pth
+            "rotary_pos_emb.inv_freq",  # chatglm
         ),
     }
 
@@ -97,6 +101,7 @@ class TensorNameMap:
             "backbone.layers.{bid}.norm",                           # mamba
             "transformer.decoder_layer.{bid}.rms_norm",             # Grok
             "transformer.blocks.{bid}.norm_attn_norm.norm_1",       # dbrx
+            "encoder.layers.{bid}.input_layernorm",                 # chatglm
         ),
 
         # Attention norm 2
@@ -117,7 +122,8 @@ class TensorNameMap:
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
             "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
-            "model.layers.{bid}.self_attn.qkv_proj"                                # phi3
+            "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
+            "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
         ),
 
         # Attention query
@@ -128,7 +134,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.q_proj",                           # gpt-j
             "model.layers.layers.{bid}.self_attn.q_proj",                # plamo
             "model.layers.{bid}.attention.wq",                           # internlm2
-            "transformer.decoder_layer.{bid}.multi_head_attention.query" # Grok
+            "transformer.decoder_layer.{bid}.multi_head_attention.query",# Grok
         ),
 
         # Attention key
@@ -140,7 +146,7 @@ class TensorNameMap:
             "transformer.h.{bid}.attn.k",                              # refact
             "model.layers.layers.{bid}.self_attn.k_proj",              # plamo
             "model.layers.{bid}.attention.wk",                         # internlm2
-            "transformer.decoder_layer.{bid}.multi_head_attention.key" # Grok
+            "transformer.decoder_layer.{bid}.multi_head_attention.key",# Grok
         ),
 
         # Attention value
@@ -175,6 +181,7 @@ class TensorNameMap:
             "encoder.layers.{bid}.attn.out_proj",                           # nomic-bert
             "transformer.decoder_layer.{bid}.multi_head_attention.linear",  # Grok
             "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj",        # dbrx
+            "encoder.layers.{bid}.self_attention.dense",                    # chatglm
         ),
 
         # Attention output norm
@@ -206,6 +213,7 @@ class TensorNameMap:
             "h.{bid}.ln_2",                                                  # gpt2
             "model.layers.{bid}.ffn_norm",                                   # internlm2
             "transformer.decoder_layer.{bid}.rms_norm_2",                    # Grok
+            "encoder.layers.{bid}.post_attention_layernorm",                 # chatglm
         ),
 
         MODEL_TENSOR.FFN_GATE_INP: (
@@ -245,6 +253,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.c_fc",                            # starcoder2
             "encoder.layer.{bid}.mlp.gated_layers_v",                 # jina-bert-v2
             "model.layers.{bid}.residual_mlp.w3",                     # arctic
+            "encoder.layers.{bid}.mlp.dense_h_to_4h",                 # chatglm
         ),
 
         MODEL_TENSOR.FFN_UP_EXP: (
@@ -311,6 +320,7 @@ class TensorNameMap:
             "model.layers.{bid}.mlp.c_proj",                          # starcoder2
             "encoder.layer.{bid}.mlp.wo",                             # jina-bert-v2
             "model.layers.{bid}.residual_mlp.w2",                     # arctic
+            "encoder.layers.{bid}.mlp.dense_4h_to_h",                 # chatglm
         ),
 
         MODEL_TENSOR.FFN_DOWN_EXP: (

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.9.0"
+version = "0.9.1"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [

llama.cpp

@@ -223,6 +223,7 @@ enum llm_arch {
     LLM_ARCH_OLMO,
     LLM_ARCH_ARCTIC,
     LLM_ARCH_DEEPSEEK2,
+    LLM_ARCH_CHATGLM,
     LLM_ARCH_UNKNOWN,
 };
 
@@ -261,6 +262,7 @@ static const std::map<llm_arch, const char *> LLM_ARCH_NAMES = {
     { LLM_ARCH_OLMO,            "olmo"         },
     { LLM_ARCH_ARCTIC,          "arctic"       },
     { LLM_ARCH_DEEPSEEK2,       "deepseek2"    },
+    { LLM_ARCH_CHATGLM,         "chatglm"      },
     { LLM_ARCH_UNKNOWN,         "(unknown)"    },
 };
 
@@ -1108,6 +1110,21 @@ static const std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NA
             { LLM_TENSOR_FFN_UP_SHEXP,       "blk.%d.ffn_up_shexp" },
         },
     },
+    {
+        LLM_ARCH_CHATGLM,
+        {
+            { LLM_TENSOR_TOKEN_EMBD,      "token_embd" },
+            { LLM_TENSOR_ROPE_FREQS,      "rope_freqs" },
+            { LLM_TENSOR_OUTPUT_NORM,     "output_norm" },
+            { LLM_TENSOR_OUTPUT,          "output" },
+            { LLM_TENSOR_ATTN_NORM,       "blk.%d.attn_norm" },
+            { LLM_TENSOR_ATTN_QKV,        "blk.%d.attn_qkv" },
+            { LLM_TENSOR_ATTN_OUT,        "blk.%d.attn_output" },
+            { LLM_TENSOR_FFN_NORM,        "blk.%d.ffn_norm" },
+            { LLM_TENSOR_FFN_UP,          "blk.%d.ffn_up" },
+            { LLM_TENSOR_FFN_DOWN,        "blk.%d.ffn_down" },
+        },
+    },
     {
         LLM_ARCH_UNKNOWN,
         {
@@ -4486,6 +4503,14 @@ static void llm_load_hparams(
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
             } break;
+        case LLM_ARCH_CHATGLM:
+        {
+            ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+            switch (hparams.n_layer) {
+                case 28: model.type = e_model::MODEL_7B; break;
+                default: model.type = e_model::MODEL_UNKNOWN;
+            }
+        } break;
         default: (void)0;
     }
 
@@ -6319,6 +6344,36 @@ static bool llm_load_tensors(
                         }
                     }
                 } break;
+            case LLM_ARCH_CHATGLM:
+                {
+                    model.tok_embd   = ml.create_tensor(ctx_input,  tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab});
+
+                    // output
+                    {
+                        model.output_norm   = ml.create_tensor(ctx_output,       tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+                        model.output        = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab});
+                    }
+
+                    for (int i = 0; i < n_layer; ++i) {
+                        ggml_context * ctx_layer = ctx_for_layer(i);
+                        ggml_context * ctx_split = ctx_for_layer_split(i);
+
+                        auto & layer = model.layers[i];
+
+                        layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+                        layer.wqkv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + (hparams.n_embd_head_k << 2)});
+                        layer.bqkv = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + (hparams.n_embd_head_k << 2)});
+
+                        layer.wo   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+                        layer.ffn_norm   = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd});
+
+                        layer.ffn_up     = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2});
+
+                        layer.ffn_down   = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
+                    }
+                } break;
             default:
                 throw std::runtime_error("unknown architecture");
         }
@@ -6553,6 +6608,7 @@ enum llm_ffn_op_type {
     LLM_FFN_GELU,
     LLM_FFN_RELU,
     LLM_FFN_RELU_SQR,
+    LLM_FFN_SWIGLU,
 };
 
 enum llm_ffn_gate_type {
@@ -6743,6 +6799,19 @@ static struct ggml_tensor * llm_build_ffn(
                 cur = ggml_sqr(ctx, cur);
                 cb(cur, "ffn_sqr(relu)", il);
             } break;
+        case LLM_FFN_SWIGLU:
+            {
+                // Project to 4h. If using swiglu double the output width, see https://arxiv.org/pdf/2002.05202.pdf
+                int64_t split_point = cur->ne[0] / 2;
+                struct ggml_tensor * x0 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], 0));
+                struct ggml_tensor * x1 = ggml_cont(ctx, ggml_view_2d(ctx, cur, split_point, cur->ne[1], cur->nb[1], split_point * ggml_element_size(cur)));
+
+                x0 = ggml_silu(ctx, x0);
+                cb(cur, "ffn_silu", il);
+
+                cur = ggml_mul(ctx, x0, x1);
+                cb(cur, "ffn_mul", il);
+            } break;
     }
 
     if (type_gate == LLM_FFN_PAR) {
@@ -11334,6 +11403,119 @@ struct llm_build_context {
         return gf;
     }
 
+    struct ggml_cgraph * build_chatglm() {
+        struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+        const int64_t n_embd_head = hparams.n_embd_head_v;
+        const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
+        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+
+        struct ggml_tensor * cur;
+        struct ggml_tensor * inpL;
+
+        inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+        // inp_pos - contains the positions
+        struct ggml_tensor * inp_pos = build_inp_pos();
+
+        // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+        struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+        for (int il = 0; il < n_layer; ++il) {
+            struct ggml_tensor * inpSA = inpL;
+
+            cur = llm_build_norm(ctx0, inpL, hparams,
+                    model.layers[il].attn_norm,
+                    NULL,
+                    LLM_NORM_RMS, cb, il);
+            cb(cur, "attn_norm", il);
+
+            // self-attention
+            {
+                struct ggml_tensor * Qcur = nullptr;
+                struct ggml_tensor * Kcur = nullptr;
+                struct ggml_tensor * Vcur = nullptr;
+
+                cur = ggml_mul_mat(ctx0, model.layers[il].wqkv, cur);
+                cb(cur, "wqkv", il);
+
+                cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
+                cb(cur, "bqkv", il);
+
+                Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd,     n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+                Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+                Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)));
+
+                cb(Qcur, "Qcur", il);
+                cb(Kcur, "Kcur", il);
+                cb(Vcur, "Vcur", il);
+ 
+                Qcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Qcur, "Qcur_rope", il);
+
+                Kcur = ggml_rope_ext(
+                    ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos, nullptr,
+                    n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+                    ext_factor, attn_factor, beta_fast, beta_slow
+                );
+                cb(Kcur, "Kcur_rope", il);
+
+                cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf,
+                        model.layers[il].wo, NULL,
+                        Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+
+            }
+
+            if (il == n_layer - 1) {
+                // skip computing output for unused tokens
+                struct ggml_tensor * inp_out_ids = build_inp_out_ids();
+                cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+                inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+            }
+
+            // Add the input
+            struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+            cb(ffn_inp, "ffn_inp", il);
+
+            // FF
+            {
+                cur = llm_build_norm(ctx0, ffn_inp, hparams,
+                        model.layers[il].ffn_norm,
+                        NULL,
+                        LLM_NORM_RMS, cb, il);
+                cb(cur, "ffn_norm", il);
+
+                cur = llm_build_ffn(ctx0, cur,
+                        model.layers[il].ffn_up,   NULL,
+                        NULL,                      NULL,
+                        model.layers[il].ffn_down, NULL,
+                        NULL,
+                        LLM_FFN_SWIGLU, LLM_FFN_SEQ, cb, il);
+                cb(cur, "ffn_out", il);
+
+            }
+
+            inpL = ggml_add(ctx0, cur, ffn_inp);
+            cb(inpL, "l_out", il);
+        }
+
+        cur = llm_build_norm(ctx0, inpL, hparams,
+                model.output_norm,
+                NULL,
+                LLM_NORM_RMS, cb, -1);
+        cb(cur, "result_norm", -1);
+
+        cur = ggml_mul_mat(ctx0, model.output, cur);
+        cb(cur, "result_output", -1);
+
+        ggml_build_forward_expand(gf, cur);
+
+        return gf;
+    }
 };
 
 static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
@@ -11556,6 +11738,10 @@ static struct ggml_cgraph * llama_build_graph(
             {
                 result = llm.build_deepseek2();
             } break;
+        case LLM_ARCH_CHATGLM:
+            {
+                result = llm.build_chatglm();
+            } break;
         default:
             GGML_ASSERT(false);
     }
@@ -16550,6 +16736,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) {
         case LLM_ARCH_OLMO:
         case LLM_ARCH_ARCTIC:
         case LLM_ARCH_DEEPSEEK2:
+        case LLM_ARCH_CHATGLM:
             return LLAMA_ROPE_TYPE_NORM;
 
         // the pairs of head values are offset by n_rot/2
@@ -18541,6 +18728,19 @@ static int32_t llama_chat_apply_template_internal(
         if (add_ass) {
             ss << "<|start_header_id|>assistant<|end_header_id|>\n\n";
         }
+    } else if (tmpl == "chatglm3" ||
+        (tmpl.find("add_generation_prompt") != std::string::npos &&
+        tmpl.find("for message in messages") != std::string::npos &&
+        tmpl.find("loop.first") != std::string::npos)) {
+        // chatglm3-6b
+        ss << "[gMASK]" << "sop";
+        for (auto message : chat) {
+            std::string role(message->role);
+            ss << "<|" << role << "|>" << "\n " << message->content;
+        }
+        if (add_ass) {
+            ss << "<|assistant|>";
+        }
     } else {
         // template not supported
         return -1;

tests/test-chat-template.cpp

@@ -56,7 +56,9 @@ int main(void) {
         //Phi-3-medium
         "{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
         //Phi-3-vision
-        "{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{- '<|assistant|>\n' -}}{% endif %}"
+        "{% for message in messages %}{{'<|' + message['role'] + '|>' + '\n' + message['content'] + '<|end|>\n' }}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{- '<|assistant|>\n' -}}{% endif %}",
+        // ChatGLM3
+        "{% for message in messages %}{% if loop.first %}[gMASK]sop<|{{ message['role'] }}|>\n {{ message['content'] }}{% else %}<|{{ message['role'] }}|>\n {{ message['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}",
     };
     std::vector<std::string> expected_output = {
         // teknium/OpenHermes-2.5-Mistral-7B
@@ -93,6 +95,8 @@ int main(void) {
         "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
         //Phi-3-vision
         "<|system|>\nYou are a helpful assistant<|end|>\n<|user|>\nHello<|end|>\n<|assistant|>\nHi there<|end|>\n<|user|>\nWho are you<|end|>\n<|assistant|>\n   I am an assistant   <|end|>\n<|user|>\nAnother question<|end|>\n<|assistant|>\n",
+        // ChatGLM3
+        "[gMASK]sop<|system|>\n You are a helpful assistant<|user|>\n Hello<|assistant|>\n Hi there<|user|>\n Who are you<|assistant|>\n    I am an assistant   <|user|>\n Another question<|assistant|>",
     };
     std::vector<char> formatted_chat(1024);
     int32_t res;