return QwenModel
if model_architecture == "MixtralForCausalLM":
return MixtralModel
+ if model_architecture == "GPT2LMHeadModel":
+ return GPT2Model
if model_architecture == "PhiForCausalLM":
return Phi2Model
if model_architecture == "PlamoForCausalLM":
return gguf.MODEL_ARCH.QWEN
if arch == "MixtralForCausalLM":
return gguf.MODEL_ARCH.LLAMA
+ if arch == "GPT2LMHeadModel":
+ return gguf.MODEL_ARCH.GPT2
if arch == "PhiForCausalLM":
return gguf.MODEL_ARCH.PHI2
if arch == "PlamoForCausalLM":
self.gguf_writer.add_tensor(new_name, data)
+class GPT2Model(Model):
+ def set_gguf_parameters(self):
+ self.gguf_writer.add_name(self.dir_model.name)
+ self.gguf_writer.add_block_count(self.hparams["n_layer"])
+ self.gguf_writer.add_context_length(self.hparams["n_ctx"])
+ self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+ self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+ self.gguf_writer.add_head_count(self.hparams["n_head"])
+ self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+ self.gguf_writer.add_file_type(self.ftype)
+
+ def write_tensors(self):
+ block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+ tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+
+ for name, data_torch in self.get_tensors():
+ # we don't need these
+ if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq", ".attn.bias")):
+ continue
+
+ if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):
+ data_torch = data_torch.transpose(1, 0)
+
+ 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()
+
+ # 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}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+ self.gguf_writer.add_tensor(new_name, data)
+
+ # note: GPT2 output is tied to (same as) wte in original model
+ if new_name == "token_embd.weight":
+ print(f"output.weight, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+ self.gguf_writer.add_tensor("output.weight", data)
+
+
class Phi2Model(Model):
def set_gguf_parameters(self):
block_count = self.hparams["n_layer"]
"tok_embeddings", # llama-pth
"embeddings.word_embeddings", # bert
"language_model.embedding.word_embeddings", # persimmon
+ "wte", # gpt2
"transformer.embd.wte", # phi2
),
MODEL_TENSOR.POS_EMBD: (
"transformer.wpe", # gpt2
"embeddings.position_embeddings", # bert
+ "wpe", # gpt2
),
# Output
"norm", # llama-pth
"embeddings.LayerNorm", # bert
"transformer.norm_f", # mpt
- "ln_f", # refact bloom qwen
+ "ln_f", # refact bloom qwen gpt2
"language_model.encoder.final_layernorm", # persimmon
"lm_head.ln", # phi2
),
"encoder.layer.{bid}.attention.output.LayerNorm", # bert
"language_model.encoder.layers.{bid}.input_layernorm", # persimmon
"model.layers.{bid}.ln1", # yi
+ "h.{bid}.ln_1", # gpt2
"transformer.h.{bid}.ln", # phi2
"model.layers.layers.{bid}.norm", # plamo
),
"transformer.h.{bid}.self_attention.query_key_value", # falcon
"h.{bid}.self_attention.query_key_value", # bloom
"language_model.encoder.layers.{bid}.self_attention.query_key_value", # persimmon
+ "h.{bid}.attn.c_attn", # gpt2
"transformer.h.{bid}.mixer.Wqkv", # phi2
),
"encoder.layer.{bid}.attention.output.dense", # bert
"transformer.h.{bid}.attn.out_proj", # gpt-j
"language_model.encoder.layers.{bid}.self_attention.dense", # persimmon
+ "h.{bid}.attn.c_proj", # gpt2
"transformer.h.{bid}.mixer.out_proj", # phi2
"model.layers.layers.{bid}.self_attn.o_proj", # plamo
),
"encoder.layer.{bid}.output.LayerNorm", # bert
"language_model.encoder.layers.{bid}.post_attention_layernorm", # persimmon
"model.layers.{bid}.ln2", # yi
+ "h.{bid}.ln_2", # gpt2
),
MODEL_TENSOR.FFN_GATE_INP: (
"transformer.h.{bid}.mlp.fc_in", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_h_to_4h", # persimmon
"transformer.h.{bid}.mlp.w1", # qwen
+ "h.{bid}.mlp.c_fc", # gpt2
"transformer.h.{bid}.mlp.fc1", # phi2
"model.layers.layers.{bid}.mlp.up_proj", # plamo
),
"encoder.layer.{bid}.output.dense", # bert
"transformer.h.{bid}.mlp.fc_out", # gpt-j
"language_model.encoder.layers.{bid}.mlp.dense_4h_to_h", # persimmon
+ "h.{bid}.mlp.c_proj", # gpt2
"transformer.h.{bid}.mlp.fc2", # phi2
"model.layers.layers.{bid}.mlp.down_proj", # plamo
),
LLM_ARCH_GPT2,
{
{ LLM_TENSOR_TOKEN_EMBD, "token_embd" },
+ { LLM_TENSOR_POS_EMBD, "position_embd" },
+ { 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" },
},
},
{
MODEL_40B,
MODEL_65B,
MODEL_70B,
+ MODEL_SMALL,
+ MODEL_MEDIUM,
+ MODEL_LARGE,
+ MODEL_XL,
};
static const size_t kiB = 1024;
static const char * llama_model_type_name(e_model type) {
switch (type) {
- case MODEL_1B: return "1B";
- case MODEL_3B: return "3B";
- case MODEL_7B: return "7B";
- case MODEL_8B: return "8B";
- case MODEL_13B: return "13B";
- case MODEL_15B: return "15B";
- case MODEL_30B: return "30B";
- case MODEL_34B: return "34B";
- case MODEL_40B: return "40B";
- case MODEL_65B: return "65B";
- case MODEL_70B: return "70B";
- default: return "?B";
+ case MODEL_1B: return "1B";
+ case MODEL_3B: return "3B";
+ case MODEL_7B: return "7B";
+ case MODEL_8B: return "8B";
+ case MODEL_13B: return "13B";
+ case MODEL_15B: return "15B";
+ case MODEL_30B: return "30B";
+ case MODEL_34B: return "34B";
+ case MODEL_40B: return "40B";
+ case MODEL_65B: return "65B";
+ case MODEL_70B: return "70B";
+ case MODEL_SMALL: return "0.1B";
+ case MODEL_MEDIUM: return "0.4B";
+ case MODEL_LARGE: return "0.8B";
+ case MODEL_XL: return "1.5B";
+ default: return "?B";
}
}
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
+ case LLM_ARCH_GPT2:
+ {
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+ switch (hparams.n_layer) {
+ case 12: model.type = e_model::MODEL_SMALL; break;
+ case 24: model.type = e_model::MODEL_MEDIUM; break;
+ case 36: model.type = e_model::MODEL_LARGE; break;
+ case 48: model.type = e_model::MODEL_XL; break;
+ default: model.type = e_model::MODEL_UNKNOWN;
+ }
+ } break;
default: (void)0;
}
layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split);
}
} break;
+ case LLM_ARCH_GPT2:
+ {
+ model.tok_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, GGML_BACKEND_CPU);
+ model.pos_embd = ml.create_tensor(ctx, tn(LLM_TENSOR_POS_EMBD, "weight"), {n_embd, hparams.n_ctx_train}, GGML_BACKEND_CPU);
+
+ // output
+ {
+ ggml_backend_type backend_norm;
+ ggml_backend_type backend_output;
+
+ if (n_gpu_layers > int(n_layer)) {
+ backend_norm = llama_backend_offload;
+ backend_output = llama_backend_offload_split;
+ } else {
+ backend_norm = GGML_BACKEND_CPU;
+ backend_output = GGML_BACKEND_CPU;
+ }
+
+ model.output_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, backend_norm);
+ model.output_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, backend_norm);
+ model.output = ml.create_tensor(ctx, tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, backend_output);
+ }
+
+ const uint32_t n_ff = hparams.n_ff;
+
+ const int i_gpu_start = n_layer - n_gpu_layers;
+
+ model.layers.resize(n_layer);
+
+ for (uint32_t i = 0; i < n_layer; ++i) {
+ const ggml_backend_type backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload; // NOLINT
+ const ggml_backend_type backend_split = int(i) < i_gpu_start ? GGML_BACKEND_CPU : llama_backend_offload_split; // NOLINT
+
+ auto & layer = model.layers[i];
+
+ layer.attn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, backend);
+ layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, backend);
+
+ layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
+ layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, backend);
+
+ layer.wo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, backend_split);
+ layer.bo = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, backend);
+
+ layer.ffn_norm = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
+ layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, backend);
+
+ layer.ffn_down = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
+ layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, backend);
+
+ layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, backend_split);
+ layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff}, backend);
+ }
+ } break;
default:
throw std::runtime_error("unknown architecture");
}
return gf;
}
+
+ struct ggml_cgraph * build_gpt2() {
+ struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+ struct ggml_tensor * cur;
+ struct ggml_tensor * pos;
+ struct ggml_tensor * inpL;
+
+ inpL = llm_build_inp_embd(ctx0, hparams, batch, model.tok_embd, cb);
+ cb(inpL, "inp_embd", -1);
+
+ // inp_pos - contains the positions
+ struct ggml_tensor * inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
+ cb(inp_pos, "inp_pos", -1);
+
+ // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+ struct ggml_tensor * KQ_mask = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_kv, n_tokens, 1);
+ cb(KQ_mask, "KQ_mask", -1);
+
+ pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos);
+ cb(pos, "pos_embd", -1);
+
+ inpL = ggml_add(ctx0, inpL, pos);
+ cb(inpL, "inpL", -1);
+
+ for (int il = 0; il < n_layer; ++il) {
+ cur = llm_build_norm(ctx0, inpL, hparams,
+ model.layers[il].attn_norm,
+ model.layers[il].attn_norm_b,
+ LLM_NORM, cb, il);
+ cb(cur, "attn_norm", il);
+
+ // self-attention
+ {
+ 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);
+
+ struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0*sizeof(float)*(n_embd)));
+ struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1*sizeof(float)*(n_embd)));
+ struct ggml_tensor * 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_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
+
+ llm_build_kv_store(ctx0, hparams, kv_self, gf, Kcur, Vcur, n_ctx, n_tokens, kv_head, cb, il);
+
+ cur = llm_build_kqv(ctx0, model, hparams, kv_self,
+ model.layers[il].wo, model.layers[il].bo,
+ Qcur, KQ_mask, n_ctx, n_tokens, n_kv, -1.0f, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+ cb(cur, "kqv_out", il);
+ }
+
+ // add the input
+ struct ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
+ cb(ffn_inp, "ffn_inp", il);
+
+ // FF
+ {
+ cur = llm_build_norm(ctx0, ffn_inp, hparams,
+ model.layers[il].ffn_norm,
+ model.layers[il].ffn_norm_b,
+ LLM_NORM, cb, il);
+ cb(cur, "ffn_norm", il);
+
+ cur = llm_build_ffn(ctx0, cur,
+ model.layers[il].ffn_up, model.layers[il].ffn_up_b,
+ NULL, NULL,
+ model.layers[il].ffn_down, model.layers[il].ffn_down_b,
+ NULL,
+ LLM_FFN_GELU, 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,
+ model.output_norm_b,
+ LLM_NORM, 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;
+ }
};
//
{
result = llm.build_plamo();
} break;
+ case LLM_ARCH_GPT2:
+ {
+ result = llm.build_gpt2();
+ } break;
default:
GGML_ASSERT(false);
}
overview / pkg / ggml / sources / llama.cpp / commitdiff