Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

llama : add phixtral support #4912

Draft
wants to merge 1 commit into
base: master
Choose a base branch
from
+107 −11
Draft

llama : add phixtral support #4912

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 11 additions & 2 deletions convert-hf-to-gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1080,10 +1080,15 @@ class Phi2Model(Model):
def set_gguf_parameters(self):
block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])

rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])

if "partial_rotary_factor" in self.hparams:
rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
n_rot = int(rot_pct * n_embd) // n_head
else:
n_rot = get_key_opts(self.hparams, ["rotary_dim", "n_rot"])

self.gguf_writer.add_name("Phi2")
self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))

Expand All @@ -1093,10 +1098,14 @@ def set_gguf_parameters(self):
self.gguf_writer.add_head_count(n_head)
self.gguf_writer.add_head_count_kv(n_head)
self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
self.gguf_writer.add_rope_dimension_count(n_rot)
self.gguf_writer.add_file_type(self.ftype)
self.gguf_writer.add_add_bos_token(False)

# phixtral
self.gguf_writer.add_expert_count(self.hparams.get("num_local_experts", 0))
self.gguf_writer.add_expert_used_count(self.hparams.get("num_experts_per_tok", 0))


class PlamoModel(Model):
def set_vocab(self):
Expand Down
3 changes: 3 additions & 0 deletions gguf-py/gguf/constants.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -393,9 +393,12 @@ class MODEL_TENSOR(IntEnum):
MODEL_TENSOR.ATTN_K,
MODEL_TENSOR.ATTN_V,
MODEL_TENSOR.ATTN_OUT,
MODEL_TENSOR.FFN_GATE_INP,
MODEL_TENSOR.FFN_NORM,
MODEL_TENSOR.FFN_DOWN,
MODEL_TENSOR.FFN_UP,
MODEL_TENSOR.FFN_DOWN_EXP,
MODEL_TENSOR.FFN_UP_EXP,
]
# TODO
}
Expand Down
3 changes: 3 additions & 0 deletions gguf-py/gguf/tensor_mapping.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -173,6 +173,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_GATE_INP: (
"layers.{bid}.feed_forward.gate", # mixtral
"model.layers.{bid}.block_sparse_moe.gate", # mixtral
"transformer.h.{bid}.moe.gate", # phixtral
),

# Feed-forward up
Expand All @@ -198,6 +199,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_UP_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w3", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w3", # mixtral
"transformer.h.{bid}.moe.mlp.{xid}.fc1", # phixtral
),

# AWQ-activation gate
Expand Down Expand Up @@ -240,6 +242,7 @@ class TensorNameMap:
MODEL_TENSOR.FFN_DOWN_EXP: (
"layers.{bid}.feed_forward.experts.{xid}.w2", # mixtral
"model.layers.{bid}.block_sparse_moe.experts.{xid}.w2", # mixtral
"transformer.h.{bid}.moe.mlp.{xid}.fc2", # phixtral
),

MODEL_TENSOR.ATTN_Q_NORM: (
Expand Down
99 changes: 90 additions & 9 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -578,8 +578,11 @@ static std::map<llm_arch, std::map<llm_tensor, std::string>> LLM_TENSOR_NAMES =
{ LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
{ LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
{ LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
{ LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" },
{ LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
{ LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
{ LLM_TENSOR_FFN_DOWN_EXP, "blk.%d.ffn_down.%d" },
{ LLM_TENSOR_FFN_UP_EXP, "blk.%d.ffn_up.%d" },
},
},
{
Expand Down Expand Up @@ -1425,16 +1428,20 @@ struct llama_layer {
struct ggml_tensor * ffn_down; // w2
struct ggml_tensor * ffn_up; // w3

// ff bias
struct ggml_tensor * ffn_down_b; // b2
struct ggml_tensor * ffn_up_b; // b3
struct ggml_tensor * ffn_act;

// ff MoE
struct ggml_tensor * ffn_gate_inp;
struct ggml_tensor * ffn_gate_exp[LLAMA_MAX_EXPERTS];
struct ggml_tensor * ffn_down_exp[LLAMA_MAX_EXPERTS];
struct ggml_tensor * ffn_up_exp [LLAMA_MAX_EXPERTS];

// ff bias
struct ggml_tensor * ffn_down_b; // b2
struct ggml_tensor * ffn_up_b; // b3
struct ggml_tensor * ffn_act;
// ff MoE bias
struct ggml_tensor * ffn_down_b_exp[LLAMA_MAX_EXPERTS];
struct ggml_tensor * ffn_up_b_exp [LLAMA_MAX_EXPERTS];
};

struct llama_kv_cell {
Expand Down Expand Up @@ -3696,11 +3703,29 @@ static bool llm_load_tensors(
layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
layer.bo = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd});

layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd});
layer.ffn_gate_inp = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd}, false);

layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
if (layer.ffn_gate_inp == nullptr) {
GGML_ASSERT(hparams.n_expert == 0);
GGML_ASSERT(hparams.n_expert_used == 0);

layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd});
layer.ffn_down_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd});

layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
layer.ffn_up_b = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ff});
} else {
GGML_ASSERT(hparams.n_expert > 0);
GGML_ASSERT(hparams.n_expert_used > 0);

for (uint32_t x = 0; x < hparams.n_expert; ++x) {
layer.ffn_down_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN_EXP, "weight", i, x), {n_ff, n_embd});
layer.ffn_down_b_exp[x] = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_DOWN_EXP, "bias", i, x), {n_embd});

layer.ffn_up_exp[x] = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP_EXP, "weight", i, x), {n_embd, n_ff});
layer.ffn_up_b_exp[x] = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_UP_EXP, "bias", i, x), {n_ff});
}
}
}
} break;
case LLM_ARCH_PLAMO:
Expand Down Expand Up @@ -5704,14 +5729,70 @@ struct llm_build_context {
}

// FF
{
if (model.layers[il].ffn_gate_inp == nullptr) {
ffn_output = llm_build_ffn(ctx0, attn_norm_output,
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(ffn_output, "ffn_out", il);
} else {
// MoE branch
ggml_tensor * logits = ggml_mul_mat(ctx0, model.layers[il].ffn_gate_inp, cur); // [n_tokens, num_experts]
cb(logits, "ffn_moe_logits", il);

ggml_tensor * probs = ggml_soft_max(ctx0, logits); // [n_tokens, num_experts]
cb(probs, "ffn_moe_probs", il);

// select experts
ggml_tensor * selected_experts = ggml_top_k(ctx0, probs, n_expert_used); // [n_tokens, num_experts_per_tok]
cb(selected_experts->src[0], "ffn_moe_argsort", il);

ggml_tensor * weights = ggml_get_rows(ctx0,
ggml_reshape_3d(ctx0, probs, 1, n_expert, n_tokens), selected_experts);
cb(weights, "ffn_moe_weights", il);

weights = ggml_reshape_2d(ctx0, weights, n_expert_used, n_tokens); // [n_tokens, num_experts_per_tok]

ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights);
cb(weights_sum, "ffn_moe_weights_sum", il);

weights = ggml_div(ctx0, weights, weights_sum); // [n_tokens, num_experts_per_tok]
cb(weights, "ffn_moe_weights_norm", il);

// compute expert outputs
ggml_tensor * moe_out = nullptr;

for (int i = 0; i < n_expert_used; ++i) {
ggml_tensor * cur_expert;

ggml_tensor * cur_up = ggml_mul_mat_id(ctx0, model.layers[il].ffn_up_exp, n_expert, selected_experts, i, cur);
#pragma message "TODO: implement ggml_add_id"
//cur_up = ggml_add_id(ctx0, cur_up, model.layers[il].ffn_up_b_exp, n_expert, selected_experts, i);
cb(cur_up, "ffn_moe_up", il);

cur_up = ggml_gelu(ctx0, cur_up);
cb(cur_up, "ffn_moe_gelu", il);

cur_expert = ggml_mul_mat_id(ctx0, model.layers[il].ffn_down_exp, n_expert, selected_experts, i, cur_up); // [n_tokens, n_embd]
#pragma message "TODO: implement ggml_add_id"
//cur_expert = ggml_add_id(ctx0, cur_expert, model.layers[il].ffn_down_b_exp, n_expert, selected_experts, i);
cb(cur_expert, "ffn_moe_down", il);

cur_expert = ggml_mul(ctx0, cur_expert,
ggml_view_2d(ctx0, weights, 1, n_tokens, weights->nb[1], i*weights->nb[0]));
cb(cur_expert, "ffn_moe_weighted", il);

if (i == 0) {
moe_out = cur_expert;
} else {
moe_out = ggml_add(ctx0, moe_out, cur_expert);
cb(moe_out, "ffn_moe_out", il);
}
}

ffn_output = moe_out;
}

cur = ggml_add(ctx0, cur, ffn_output);
Expand Down
Loading