model: Add support for SenseNova U1 model (#28919)

* model : add SenseNova U1 understanding support

Add support for the dense SenseNova U1 understanding branch, including
the U1.5-8B-MoT model.

- Add HF-to-GGUF conversion for NEOChatModel, excluding MoT generation
  tensors and concatenating temporal and spatial Q/K normalization
  weights.
- Register the sensenova_u1 GGUF architecture and custom spatial RoPE
  metadata.
- Implement temporal/spatial Q/K normalization with 4-axis RoPE.
- Keep U1-specific text position handling in the model implementation.
- Add synthetic architecture coverage for sensenova_u1.

Only the dense understanding branch is supported; generation
branches are intentionally rejected during conversion.

* conversion : treat one-expert SenseNova U1 configs as dense

* mtmd : add SenseNova U1 vision support

Add conversion support for the SenseNova U1 vision encoder and implement
its CPU mtmd projector graph.

The graph applies patch embedding, GELU, adjacent-pair 2D RoPE, spatial
downsampling, and dense projection for dynamic-resolution image inputs.

* mtmd : fix SenseNova U1 vision token layout

* model : clean up SenseNova U1 RoPE metadata

Replace the hard-coded spatial RoPE key with shared GGUF and LLM_KV
definitions used by conversion, model loading, and architecture tests.

Clarify why SenseNova U1 retains its custom text-position input: U1 uses
[t, 0, 0, 0], while the generic M-RoPE input uses [t, t, t, 0].

* mtmd : simplify SenseNova U1 projector integration

Map the SenseNova U1 dense projection to the standard mm.0 weight and
bias tensors and reuse the shared projector fields in the loader and graph.

Read vision parameters from the checkpoint configuration and remove the
U1-specific tensor names and model fields while preserving its strided
projection, bias, and row-major token layout.

* mtmd : fix SenseNova U1 image positions and attention

* fix spacing

* mtmd : restore generic test input chunks

* mtmd : restore SenseNova U1 integration after upstream merge

* Update conversion/sensenova_u1.py

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update src/models/sensenova-u1.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* model : remove SenseNova U1 BF16 activation casts

Use standard backend precision handling for QKV, FFN, and output
projections instead of casting inputs based on the output tensor type.

Assisted-by: Codex

* Update src/models/sensenova-u1.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
This commit is contained in:
Maphist0andSigbjørn Skjæret authored and GitHub committed 2026-10-21 21:54:16 +02:00
1 parent d454a204ef
commit 7775ffb7e4
18 files changed
+470 -2

No files matched your search

+190
View File
@@ -0,0 +1,190 @@
#include "models.h"
namespace {
class llm_graph_input_pos_sensenova_u1 : public llm_graph_input_pos {
public:
llm_graph_input_pos_sensenova_u1() : llm_graph_input_pos(4) {}
void set_input(const llama_ubatch * ubatch) override {
if (ubatch->pos && pos) {
const int64_t n_tokens = ubatch->n_tokens;
// U1 expects text-token positions in the temporal plane only: [t, 0, 0, 0].
// The generic M-RoPE input uses [t, t, t, 0], so U1 needs a custom input class.
if (ubatch->token) {
std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd, 0);
for (int i = 0; i < n_tokens; ++i) {
pos_data[i] = ubatch->pos[i];
}
ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
} else {
ggml_backend_tensor_set(pos, ubatch->pos, 0, n_tokens*n_pos_per_embd*ggml_element_size(pos));
}
}
}
};
}
void llama_model_sensenova_u1::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
ml.get_key(LLM_KV_ROPE_FREQ_BASE_SPATIAL, rope_freq_base_spatial);
if (hparams.n_embd_head_k() % 4 != 0 || hparams.n_embd_head_v() != hparams.n_embd_head_k()) {
throw std::runtime_error("SenseNova U1 requires equal Q/K/V head sizes divisible by four");
}
switch (hparams.n_layer()) {
case 42: type = LLM_TYPE_8B; break;
default: type = LLM_TYPE_UNKNOWN;
}
}
void llama_model_sensenova_u1::load_arch_tensors(llama_model_loader &) {
LLAMA_LOAD_LOCALS;
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
// output
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
// if output is NULL, init from the input tok embed
if (output == NULL) {
output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
}
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0);
layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
}
}
std::unique_ptr<llm_graph_context> llama_model_sensenova_u1::build_arch_graph(const llm_graph_params & params) const {
return std::make_unique<graph>(*this, params);
}
llama_model_sensenova_u1::graph::graph(const llama_model_sensenova_u1 & model, const llm_graph_params & params) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v();
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
ggml_tensor * cur;
ggml_tensor * inpL;
auto normalize = [&](ggml_tensor * x, ggml_tensor * weight) {
x = ggml_rms_norm(ctx0, x, hparams.f_norm_rms_eps);
return ggml_mul(ctx0, x, weight);
};
inpL = build_inp_embd(model.tok_embd);
ggml_tensor * pos[3]{};
auto inp = std::make_unique<llm_graph_input_pos_sensenova_u1>();
auto & inp_pos = inp->pos;
inp_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, (int64_t)n_tokens*hparams.n_pos_per_embd());
ggml_set_input(inp_pos);
res->add_input(std::move(inp));
for (int axis = 0; axis < 3; ++axis) {
pos[axis] = ggml_view_1d(ctx0, inp_pos, n_tokens, axis*n_tokens*sizeof(int32_t));
}
auto normalize_and_rotate = [&](ggml_tensor * x, ggml_tensor * weight) {
const int64_t heads = x->ne[1];
x = ggml_reshape_4d(ctx0, x, n_embd_head/2, 2, heads, n_tokens);
x = normalize(x, ggml_reshape_2d(ctx0, weight, n_embd_head/2, 2));
x = ggml_reshape_3d(ctx0, x, n_embd_head, heads, n_tokens);
// Normalize temporal and spatial halves separately; rotate time, height, then width.
for (int axis = 0; axis < 3; ++axis) {
const int dims = axis == 0 ? n_embd_head/2 : n_embd_head/4;
const int offset = axis == 0 ? 0 : n_embd_head/2 + (axis - 1)*n_embd_head/4;
const float base = axis == 0 ? freq_base : model.rope_freq_base_spatial;
x = ggml_rope_ext(ctx0, x, pos[axis], nullptr, dims, GGML_ROPE_TYPE_NEOX,
n_ctx_orig, base, 1.0f, 0.0f, 1.0f, 32.0f, 1.0f);
ggml_rope_set_offset(x, offset);
}
return x;
};
auto * inp_attn = build_attn_inp_kv();
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
res->t_layer_inp[il] = inpL;
ggml_tensor * inpSA = inpL;
// norm
cur = normalize(inpL, model.layers[il].attn_norm);
cb(cur, "attn_norm", il);
// self-attention
{
// compute Q and K and RoPE them
auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
n_embd_head, n_head, n_head_kv, il);
Qcur = normalize_and_rotate(Qcur, model.layers[il].attn_q_norm);
Kcur = normalize_and_rotate(Kcur, model.layers[il].attn_k_norm);
cb(Qcur, "Qcur", il);
cb(Kcur, "Kcur", il);
cb(Vcur, "Vcur", il);
cur = build_attn(inp_attn,
model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
}
if (il == n_layer - 1 && inp_out_ids) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// feed-forward network
cur = normalize(ffn_inp, model.layers[il].ffn_norm);
cb(cur, "ffn_norm", il);
cur = build_ffn(cur,
model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_s,
model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_s,
model.layers[il].ffn_down, NULL, model.layers[il].ffn_down_s,
NULL,
LLM_FFN_SILU, LLM_FFN_PAR, il);
cb(cur, "ffn_out", il);
cur = ggml_add(ctx0, cur, ffn_inp);
cur = build_cvec(cur, il);
cb(cur, "l_out", il);
// input for next layer
inpL = cur;
}
cur = inpL;
cur = normalize(cur, model.output_norm);
cb(cur, "result_norm", -1);
res->t_embd = cur;
// lm_head
cur = build_lora_mm(model.output, cur, model.output_s);
cb(cur, "result_output", -1);
res->t_logits = cur;
}