mikhail/llama.cpp-mtp-turboquant
1
0
Fork 0
Code Issues Pull Requests Actions 33 Packages Projects Releases Wiki Activity

llama : enable chunked fused GDN path (#20340)

Browse Source 
* llama : enable chunked fused GDN path

* models : avoid Q and K repeats when using fused GDA

* cont : fix comment

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

* cont : fix the fix

Co-authored-by: Aman Gupta <amangupta052@gmail.com>

* cont : fix

* metal : add GDN kernel (#20361)

* metal : add Metal backend for GGML_OP_GATED_DELTA_NET

Add a fused Metal kernel for the gated delta net recurrence op
(#19504), enabling GPU-accelerated inference for DeltaNet-based
models (Qwen3.5, etc.) on Apple Silicon.

Supports both GDA (scalar gate) and KDA (per-row gate) modes
with head_size 64 and 128. Unsupported configurations (head_size
32, non-contiguous tensors) gracefully fall back to CPU.

Performance: Qwen3.5-0.8B Q4_K_M on M4 Max
  tg128: 170 -> 213 t/s (+25%)

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* metal : validate contiguity of all input tensors in supports_op

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* metal : add algorithm equivalence comment for GDA decay path

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* cont : unslop + optimize

* cont : clean-up

---------

Co-authored-by: Paul Flynn <paul@arkavo.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

* CUDA: AR gated delta net improvements (#20391)

* Add FastDiv to gated_delta_net_cuda

* Shard columns across warps

This reduces register pressure (avoids spill for S_v = 128) and gives
the warp-scheduler more CTAs to schedule (thus hiding data-access
latencies).

* Remove unneded include in gated_delta_net.cu

* Improve comments

* Apply code-formating

* Make sharding HIP-compatible

1. Use ggml_cuda_get_physical_warp_size() to determine warp size flexibly
2. Add test with partial warp to test sum reduction on CUDA

* Remove fastdiv_s64, as we can treat neqk1 and rq3 as uint32_t

* Rename variables

* Enable GDN also for prefill, move TODO for chunked_GDN

* Actually remove the TODO from 206890897546bd16602c3b79394fd5ea09ef199f

* Get warp size at runtime

warp_size is not known at compile time in hip host code.

* Don't expose ggml_cuda_get_physical_warp_size on host

---------

Co-authored-by: uvos <devnull@uvos.xyz>

* llama : refactor llm_build_delta_net_base API

---------

Co-authored-by: Aman Gupta <amangupta052@gmail.com>
Co-authored-by: Paul Flynn <paul@arkavo.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: Oliver Simons <osimons@nvidia.com>
Co-authored-by: uvos <devnull@uvos.xyz>
This commit is contained in:
Georgi Gerganov
2026-03-11 22:46:40 +02:00
committed by GitHub
parent f90bd1dd84
commit d28961d81e
20 changed files with 674 additions and 165 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/delta-net-base.cpp
+75 -27
View File
@@ -41,13 +41,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
if (cparams.fused_gdn_ch) {
//ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
//cb(result, LLAMA_TENSOR_NAME_FGDNCH, il);
GGML_ABORT("not implemented yet");
}
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
@@ -325,26 +318,6 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
if (cparams.fused_gdn_ar) {
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
cb(result, LLAMA_TENSOR_NAME_FGDNAR, il);
ggml_tensor * output = ggml_view_4d(ctx0, result,
S_v, H_v, n_tokens, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
ggml_tensor * new_state = ggml_view_4d(ctx0, result,
S_v, S_v, H_v, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * S_v),
ggml_row_size(result->type, S_v * S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
return {output, new_state};
}
const float scale = 1.0f / sqrtf(S_k);
q = ggml_scale(ctx0, q, scale);
@@ -401,3 +374,78 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
return {o, s};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net_fused(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t S_k = q->ne[0];
const int64_t H_k = q->ne[1];
const int64_t n_tokens = q->ne[2];
const int64_t n_seqs = q->ne[3];
const int64_t S_v = v->ne[0];
const int64_t H_v = v->ne[1];
GGML_ASSERT(S_k == S_v);
GGML_ASSERT(H_v % H_k == 0);
GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs);
GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs);
GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs);
GGML_ASSERT(g->ne[0] == 1 || g->ne[0] == S_v);
GGML_ASSERT( g->ne[1] == H_v && g->ne[2] == n_tokens && g->ne[3] == n_seqs);
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s);
if (n_tokens == 1) {
cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
} else {
cb(result, LLAMA_TENSOR_NAME_FGDN_CH, il);
}
ggml_tensor * output = ggml_view_4d(ctx0, result,
S_v, H_v, n_tokens, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens), 0);
ggml_tensor * new_state = ggml_view_4d(ctx0, result,
S_v, S_v, H_v, n_seqs,
ggml_row_size(result->type, S_v),
ggml_row_size(result->type, S_v * S_v),
ggml_row_size(result->type, S_v * S_v * H_v),
ggml_row_size(result->type, S_v * H_v * n_tokens * n_seqs));
return {output, new_state};
}
std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_net(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il) {
const int64_t n_seq_tokens = q->ne[2];
if (n_seq_tokens == 1) {
if (cparams.fused_gdn_ar) {
return build_delta_net_fused(q, k, v, g, b, s, il);
}
return build_delta_net_autoregressive(q, k, v, g, b, s, il);
}
if (cparams.fused_gdn_ch) {
return build_delta_net_fused(q, k, v, g, b, s, il);
}
return build_delta_net_chunking(q, k, v, g, b, s, il);
}
src/models/kimi-linear.cpp
+1 -3
View File
@@ -169,9 +169,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
Kcur = ggml_l2_norm(ctx0, Kcur, eps_norm);
// Choose between build_delta_net_chunking and build_delta_net_recurrent based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out = n_seq_tokens == 1 ?
build_delta_net_autoregressive(Qcur, Kcur, Vcur, g1, beta, state, il) :
build_delta_net_chunking(Qcur, Kcur, Vcur, g1, beta, state, il);
auto attn_out = build_delta_net(Qcur, Kcur, Vcur, g1, beta, state, il);
ggml_tensor * output = ggml_cont(ctx0, attn_out.first);
ggml_tensor * new_state = attn_out.second;
src/models/models.h
+20
View File
@@ -44,6 +44,26 @@ struct llm_build_delta_net_base : public llm_graph_context {
ggml_tensor * b,
ggml_tensor * s,
int il);
// use the ggml_gated_delta_net fused operator
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_fused(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il);
// choose one of two implementations above based on the number of tokens
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net(
ggml_tensor * q,
ggml_tensor * k,
ggml_tensor * v,
ggml_tensor * g,
ggml_tensor * b,
ggml_tensor * s,
int il);
};
struct llm_build_rwkv6_base : public llm_graph_context {
src/models/qwen35.cpp
+4 -8
View File
@@ -321,9 +321,9 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
if (num_k_heads != num_v_heads) {
// note: need explicit repeat only if we are not using the fused GDN
if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
// TODO: try to avoid these explicit repeats by utilizing op broadcast
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
@@ -332,12 +332,8 @@ ggml_tensor * llm_build_qwen35::build_layer_attn_linear(
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
std::pair<ggml_tensor *, ggml_tensor *> attn_out;
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
src/models/qwen35moe.cpp
+4 -8
View File
@@ -321,9 +321,9 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
if (num_k_heads != num_v_heads) {
// note: need explicit repeat only if we are not using the fused GDN
if (num_k_heads != num_v_heads && (!cparams.fused_gdn_ar || !cparams.fused_gdn_ch)) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
// TODO: try to avoid these explicit repeats by utilizing op broadcast
q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs);
}
@@ -332,12 +332,8 @@ ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear(
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
std::pair<ggml_tensor *, ggml_tensor *> attn_out;
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);
src/models/qwen3next.cpp
+3 -7
View File
@@ -406,6 +406,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
//v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs);
// if head keys and value keys are different, repeat to force tensors into matching shapes
// TODO: avoid repeats for fused GDN, needs broadcast configuration for GDN op [TAG_GGML_GDN_BCAST]
if (num_k_heads != num_v_heads) {
GGML_ASSERT(num_v_heads % num_k_heads == 0);
int64_t repeat_factor = num_v_heads / num_k_heads;
@@ -431,13 +432,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear(
cb(k_conv, "k_conv_predelta", il);
cb(v_conv, "v_conv_predelta", il);
// Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens
std::pair<ggml_tensor *, ggml_tensor *> attn_out; // pair of (output, new_state)
if (n_seq_tokens == 1) {
attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il);
} else {
attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il);
}
auto attn_out = build_delta_net(q_conv, k_conv, v_conv, gate, beta, state, il);
ggml_tensor * output = attn_out.first;
ggml_tensor * new_state = attn_out.second;
cb(output, "attn_output", il);