Skip to content

ggml: add Q4_1 support to Hexagon backend #61

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

Open
max-krasnyansky wants to merge 2 commits into
base: master
Choose a base branch
from
Open

ggml: add Q4_1 support to Hexagon backend #61

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
b5b18e8
ggml: Add Q4_1 support to Hexagon backend
google-labs-jules[bot] May 24, 2026
c71c56d
ggml: Add Q4_1 support to Hexagon backend
google-labs-jules[bot] May 24, 2026
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
319 changes: 319 additions & 0 deletions ggml/src/ggml-hexagon/ggml-hexagon.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -345,6 +345,307 @@ static void dump_packed_block_q4x4x2(const uint8_t * v, unsigned int i, size_t k
GGML_FP16_TO_FP32(d[4]), GGML_FP16_TO_FP32(d[5]), GGML_FP16_TO_FP32(d[6]), GGML_FP16_TO_FP32(d[7]));
}


static void unpack_q4_1_quants(uint8_t * qs, const block_q4_1 * x, int bi) {
static const int qk = QK4_1; // 32
for (int i = 0; i < qk / 2; i++) {
const uint8_t x0 = x->qs[i];
qs[bi * qk + i] = x0 & 0x0F;
qs[bi * qk + i + (qk / 2)] = x0 >> 4;
}

static void pack_q4_1_quants(block_q4_1 * x, const uint8_t * qs, int bi) {
static const int qk = QK4_1; // 32
for (int i = 0; i < qk / 2; i++) {
const uint8_t x0 = qs[bi * qk + i];
const uint8_t x1 = qs[bi * qk + i + qk / 2];
x->qs[i] = x0 | (x1 << 4);
}
}
}



static void repack_row_q4_1x4x2(uint8_t * y, const block_q4_1 * x, int64_t k) {
static const int qk = QK_Q4_1x4x2;
const int nb = (k + qk - 1) / qk; // number of blocks (padded)
const int nloe = k % qk; // leftovers

const int dblk_size = 8 * 2 * 2; // 8x __fp16 for d, 8x __fp16 for m. total 32 bytes
const int qblk_size = qk / 2; // int4
const int qrow_size = k / 2; // int4 (not padded to blocks)

uint8_t * y_q = y + 0; // quants first
uint8_t * y_d = y + qrow_size; // then scales & mins

// Repack the quants
for (int i = 0; i < nb; i++) {
uint8_t qs[QK_Q4_1x4x2]; // unpacked quants
unpack_q4_1_quants(qs, &x[i * 8 + 0], 0);
unpack_q4_1_quants(qs, &x[i * 8 + 1], 1);
unpack_q4_1_quants(qs, &x[i * 8 + 2], 2);
unpack_q4_1_quants(qs, &x[i * 8 + 3], 3);
unpack_q4_1_quants(qs, &x[i * 8 + 4], 4);
unpack_q4_1_quants(qs, &x[i * 8 + 5], 5);
unpack_q4_1_quants(qs, &x[i * 8 + 6], 6);
unpack_q4_1_quants(qs, &x[i * 8 + 7], 7);

bool partial = (nloe && i == nb-1);

uint8_t * q = y_q + (i * qblk_size);
for (int j = 0; j < qk / 2; j++) {
// Note: interleave the 128 elements the same way as q4_0
q[j] = partial ? (qs[j*2+1] << 4) | qs[j*2+0] : (qs[j+128] << 4) | qs[j+000];
}
}

// Repack the scales and mins
for (int i = 0; i < nb; i++) {
ggml_half * dm = (ggml_half *) (y_d + i * dblk_size);
dm[0] = x[i * 8 + 0].d;
dm[1] = x[i * 8 + 1].d;
dm[2] = x[i * 8 + 2].d;
dm[3] = x[i * 8 + 3].d;
dm[4] = x[i * 8 + 4].d;
dm[5] = x[i * 8 + 5].d;
dm[6] = x[i * 8 + 6].d;
dm[7] = x[i * 8 + 7].d;

dm[8] = x[i * 8 + 0].m;
dm[9] = x[i * 8 + 1].m;
dm[10] = x[i * 8 + 2].m;
dm[11] = x[i * 8 + 3].m;
dm[12] = x[i * 8 + 4].m;
dm[13] = x[i * 8 + 5].m;
dm[14] = x[i * 8 + 6].m;
dm[15] = x[i * 8 + 7].m;
}
}

static void unpack_row_q4_1x4x2(block_q4_1 * x, const uint8_t * y, int64_t k) {
static const int qk = QK_Q4_1x4x2;
const int nb = (k + qk - 1) / qk; // number of blocks (padded)
const int nloe = k % qk; // leftovers

const int dblk_size = 8 * 2 * 2; // 32 bytes (d and m)
const int qblk_size = qk / 2; // int4
const int qrow_size = k / 2; // int4 (not padded to blocks)

const uint8_t * y_q = y + 0;
const uint8_t * y_d = y + qrow_size;

// Unpack the quants
for (int i = 0; i < nb; i++) {
uint8_t qs[QK_Q4_1x4x2]; // unpacked quants

bool partial = (nloe && i == nb-1);

const uint8_t * q = y_q + (i * qblk_size);
for (int j = 0; j < qk / 2; j++) {
if (partial) {
qs[j*2+0] = q[j] & 0x0F;
qs[j*2+1] = q[j] >> 4;
} else {
qs[j+000] = q[j] & 0x0F;
qs[j+128] = q[j] >> 4;
}
}

pack_q4_1_quants(&x[i * 8 + 0], qs, 0);
pack_q4_1_quants(&x[i * 8 + 1], qs, 1);
pack_q4_1_quants(&x[i * 8 + 2], qs, 2);
pack_q4_1_quants(&x[i * 8 + 3], qs, 3);
pack_q4_1_quants(&x[i * 8 + 4], qs, 4);
pack_q4_1_quants(&x[i * 8 + 5], qs, 5);
pack_q4_1_quants(&x[i * 8 + 6], qs, 6);
pack_q4_1_quants(&x[i * 8 + 7], qs, 7);
}

// Unpack the scales and mins
for (int i = 0; i < nb; i++) {
const ggml_half * dm = (const ggml_half *) (y_d + i * dblk_size);
x[i * 8 + 0].d = dm[0];
x[i * 8 + 1].d = dm[1];
x[i * 8 + 2].d = dm[2];
x[i * 8 + 3].d = dm[3];
x[i * 8 + 4].d = dm[4];
x[i * 8 + 5].d = dm[5];
x[i * 8 + 6].d = dm[6];
x[i * 8 + 7].d = dm[7];

x[i * 8 + 0].m = dm[8];
x[i * 8 + 1].m = dm[9];
x[i * 8 + 2].m = dm[10];
x[i * 8 + 3].m = dm[11];
x[i * 8 + 4].m = dm[12];
x[i * 8 + 5].m = dm[13];
x[i * 8 + 6].m = dm[14];
x[i * 8 + 7].m = dm[15];
}
}

static void init_row_q4_1x4x2(block_q4_1 * x, int64_t k) {
static const int qk = QK_Q4_1x4x2;
const int nb = (k + qk - 1) / qk; // number of blocks (padded)

// Init the quants such that they unpack into zeros
uint8_t qs[QK_Q4_1x4x2]; // unpacked quants
memset(qs, 8, sizeof(qs)); // Same offset logic as q4_0 if 0 becomes 8 in q4 format, actually 0 is just 0 for q4_1 too?
// Wait, is q4_1 asymmetric or symmetric? In q4_0, value = (q - 8) * d. In q4_1, value = q * d + m.
// So to get 0, if m=0 and d=0, q can be 0. Let's just use memset 0.
// Wait, q4_0 sets qs to 8. q4_1 usually doesn't have an implicit offset. Let's look at init_row_q4_0.
// Actually, it doesn't matter much as long as it's deterministic. I will set to 0.

for (int i = 0; i < nb; i++) {
pack_q4_1_quants(&x[i * 8 + 0], qs, 0);
pack_q4_1_quants(&x[i * 8 + 1], qs, 1);
pack_q4_1_quants(&x[i * 8 + 2], qs, 2);
pack_q4_1_quants(&x[i * 8 + 3], qs, 3);
pack_q4_1_quants(&x[i * 8 + 4], qs, 4);
pack_q4_1_quants(&x[i * 8 + 5], qs, 5);
pack_q4_1_quants(&x[i * 8 + 6], qs, 6);
pack_q4_1_quants(&x[i * 8 + 7], qs, 7);
}

// Init the scales and mins
for (int i = 0; i < nb; i++) {
x[i * 8 + 0].d = 0;
x[i * 8 + 1].d = 0;
x[i * 8 + 2].d = 0;
x[i * 8 + 3].d = 0;
x[i * 8 + 4].d = 0;
x[i * 8 + 5].d = 0;
x[i * 8 + 6].d = 0;
x[i * 8 + 7].d = 0;

x[i * 8 + 0].m = 0;
x[i * 8 + 1].m = 0;
x[i * 8 + 2].m = 0;
x[i * 8 + 3].m = 0;
x[i * 8 + 4].m = 0;
x[i * 8 + 5].m = 0;
x[i * 8 + 6].m = 0;
x[i * 8 + 7].m = 0;
}
}

// repack q4_1 data into q4_1x4x2 tensor
static void repack_q4_1_q4_1x4x2(ggml_tensor * t, const void * data, size_t size) {
int64_t nrows = ggml_nrows(t);

size_t row_size = ggml_row_size(t->type, t->ne[0]);
size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_1x4x2)); // extra elements for the pad
size_t row_size_rp = row_size * 2; // extra space for tmp pad (if any)

// Ensure we don't try to read more data than is available in the source buffer 'data'
// or write more than the tensor can hold.
const size_t total_tensor_size = (size_t)nrows * row_size;
const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;

// Calculate how many full rows and how many remaining bytes we need to process.
const int64_t n_full_rows = n_bytes_to_copy / row_size;
const size_t n_rem_bytes = n_bytes_to_copy % row_size;

void * buf_pd = ggml_aligned_malloc(row_size_pd);
GGML_ASSERT(buf_pd != NULL);

void * buf_rp = ggml_aligned_malloc(row_size_rp);
GGML_ASSERT(buf_rp != NULL);

HEX_VERBOSE("ggml-hex: repack-q4_1-q4_1x4x2 %s : data %p size %zu dims %ldx%ld row-size %zu\n", t->name, data, size,
t->ne[0], nrows, row_size);

init_row_q4_1x4x2((block_q4_1 *) buf_pd, t->ne[0]); // init padded buffer to make sure the tail is all zeros

// 1. Process all the full rows
for (int64_t i = 0; i < n_full_rows; i++) {
const uint8_t * src = (const uint8_t *) data + (i * row_size);
uint8_t * dst = (uint8_t *) t->data + (i * row_size);

memcpy(buf_pd, src, row_size);
repack_row_q4_1x4x2((uint8_t *) buf_rp, (const block_q4_1 *) buf_pd, t->ne[0]);
memcpy(dst, buf_rp, row_size);
}

// 2. Process the final, potentially partial, row
if (n_rem_bytes > 0) {
const int64_t i = n_full_rows;
const uint8_t * src = (const uint8_t *) data + (i * row_size);
uint8_t * dst = (uint8_t *) t->data + (i * row_size);

// re-init the row because we are potentially copying a partial row
init_row_q4_1x4x2((block_q4_1 *) buf_pd, t->ne[0]);

// Copy only the remaining bytes from the source.
memcpy(buf_pd, src, n_rem_bytes);

// Repack the entire buffer
repack_row_q4_1x4x2((uint8_t *) buf_rp, (const block_q4_1 *) buf_pd, t->ne[0]);

// Write only the corresponding remaining bytes to the destination tensor.
memcpy(dst, buf_rp, n_rem_bytes);
}

ggml_aligned_free(buf_pd, row_size_pd);
ggml_aligned_free(buf_rp, row_size_rp);
}

// repack q4_1x4x2 tensor into q4_1 data
static void repack_q4_1x4x2_q4_1(void * data, const ggml_tensor * t, size_t size) {
int64_t nrows = ggml_nrows(t);

size_t row_size = ggml_row_size(t->type, t->ne[0]);
size_t row_size_pd = ggml_row_size(t->type, hex_round_up(t->ne[0], QK_Q4_1x4x2)); // extra elements for the pad
size_t row_size_rp = row_size * 2; // extra space for tmp pad (if any)

// Ensure we don't try to copy more data than the tensor actually contains.
const size_t total_tensor_size = (size_t)nrows * row_size;
const size_t n_bytes_to_copy = size < total_tensor_size ? size : total_tensor_size;

// Calculate how many full rows and how many remaining bytes we need to process.
const int64_t n_full_rows = n_bytes_to_copy / row_size;
const size_t n_rem_bytes = n_bytes_to_copy % row_size;

void * buf_pd = ggml_aligned_malloc(row_size_pd);
GGML_ASSERT(buf_pd != NULL);

void * buf_rp = ggml_aligned_malloc(row_size_rp);
GGML_ASSERT(buf_rp != NULL);

HEX_VERBOSE("ggml-hex: repack-q4_1x4x2-q4_1 %s : data %p size %zu dims %ldx%ld row-size %zu\n", t->name, data, size,
t->ne[0], nrows, row_size);

memset(buf_pd, 0, row_size_pd); // clear-out padded buffer to make sure the tail is all zeros

// 1. Process all the full rows
for (int64_t i = 0; i < n_full_rows; i++) {
const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
uint8_t * dst = (uint8_t *) data + (i * row_size);

memcpy(buf_rp, src, row_size);
unpack_row_q4_1x4x2((block_q4_1 *) buf_pd, (const uint8_t *) buf_rp, t->ne[0]);
memcpy(dst, buf_pd, row_size);
}

// 2. Process the final, potentially partial, row
if (n_rem_bytes > 0) {
const int64_t i = n_full_rows;
const uint8_t * src = (const uint8_t *) t->data + (i * row_size);
uint8_t * dst = (uint8_t *) data + (i * row_size);

// Copy only the remaining bytes from the source.
memcpy(buf_rp, src, n_rem_bytes);

// Unpack the entire buffer
unpack_row_q4_1x4x2((block_q4_1 *) buf_pd, (const uint8_t *) buf_rp, t->ne[0]);

// Write only the corresponding remaining bytes to the destination data buffer.
memcpy(dst, buf_pd, n_rem_bytes);
}

ggml_aligned_free(buf_pd, row_size_pd);
ggml_aligned_free(buf_rp, row_size_rp);
}
static void unpack_q4_0_quants(uint8_t * qs, const block_q4_0 * x, unsigned int bi) {
static const int qk = QK4_0;

Expand Down Expand Up @@ -1365,6 +1666,12 @@ static void ggml_backend_hexagon_buffer_set_tensor(ggml_backend_buffer_t buffer,
repack_q4_0_q4x4x2(tensor, data, size);
break;

case GGML_TYPE_Q4_1:
GGML_ASSERT(offset == 0);
GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
repack_q4_1_q4_1x4x2(tensor, data, size);
break;

case GGML_TYPE_Q8_0:
GGML_ASSERT(offset == 0);
GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
Expand Down Expand Up @@ -1407,6 +1714,12 @@ static void ggml_backend_hexagon_buffer_get_tensor(ggml_backend_buffer_t buffer,
repack_q4x4x2_q4_0(data, tensor, size);
break;

case GGML_TYPE_Q4_1:
GGML_ASSERT(offset == 0);
GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
repack_q4_1x4x2_q4_1(data, tensor, size);
break;

case GGML_TYPE_Q8_0:
GGML_ASSERT(offset == 0);
GGML_ASSERT(offset + size <= ggml_nbytes(tensor));
Expand Down Expand Up @@ -2327,6 +2640,7 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s

switch (src0->type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_MXFP4:
Expand Down Expand Up @@ -2377,6 +2691,7 @@ static bool ggml_hexagon_supported_mul_mat_id(const struct ggml_hexagon_session

switch (src0->type) {
case GGML_TYPE_Q4_0:
case GGML_TYPE_Q4_1:
case GGML_TYPE_Q8_0:
case GGML_TYPE_IQ4_NL:
case GGML_TYPE_MXFP4:
Expand Down Expand Up @@ -3598,8 +3913,12 @@ static void ggml_hexagon_init(ggml_backend_reg * reg) {
// Basic sanity checks to make sure definitions match
static_assert((unsigned int) HTP_TYPE_Q4_0 == (unsigned int) GGML_TYPE_Q4_0,
"please update hexagon_type to match ggml_type");
static_assert((unsigned int) HTP_TYPE_Q4_1 == (unsigned int) GGML_TYPE_Q4_1,
"please update hexagon_type to match ggml_type");
static_assert((unsigned int) HTP_TYPE_Q8_0 == (unsigned int) GGML_TYPE_Q8_0,
"please update hexagon_type to match ggml_type");
static_assert((unsigned int) HTP_TYPE_Q8_1 == (unsigned int) GGML_TYPE_Q8_1,
"please update hexagon_type to match ggml_type");
static_assert((unsigned int) HTP_TYPE_MXFP4 == (unsigned int) GGML_TYPE_MXFP4,
"please update hexagon_type to match ggml_type");
static_assert((unsigned int) HTP_TYPE_IQ4_NL == (unsigned int) GGML_TYPE_IQ4_NL,
Expand Down
Loading
Loading