enne2/qwentts.cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #1 from smashingtags/master

Browse Source 
cmake: add Intel Arc GPU support via SYCL backend
This commit is contained in:
Pascal
2026-05-20 14:10:11 +02:00
committed by GitHub
parent 55b7234d21 dc8d2d59a5
commit 85ad801950
3 changed files with 108 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
CMakeLists.txt
+10 -3
View File
@@ -40,8 +40,10 @@ set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
# Audio tokenizer tensor names can exceed default GGML_MAX_NAME of 64 # Audio tokenizer tensor names can exceed default GGML_MAX_NAME of 64
add_compile_definitions(GGML_MAX_NAME=128) add_compile_definitions(GGML_MAX_NAME=128)
# Harden: mark fread/fwrite/etc with warn_unused_result on all platforms # Harden: mark fread/fwrite/etc with warn_unused_result on all platforms.
if(NOT MSVC) # Disabled for SYCL: _FORTIFY_SOURCE replaces memcpy with __memcpy_chk
# which is unresolvable in GPU device code and aborts kernel compilation.
if(NOT MSVC AND NOT GGML_SYCL)
add_compile_definitions(_FORTIFY_SOURCE=2) add_compile_definitions(_FORTIFY_SOURCE=2)
endif() endif()
@@ -79,7 +81,7 @@ macro(link_ggml_backends target)
if(TARGET ggml-base) if(TARGET ggml-base)
target_link_libraries(${target} PRIVATE ggml-base) target_link_libraries(${target} PRIVATE ggml-base)
endif() endif()
foreach(backend cpu blas cuda metal vulkan) foreach(backend cpu blas cuda metal vulkan sycl)
if(TARGET ggml-${backend}) if(TARGET ggml-${backend})
get_target_property(CURRENT_BACKEND_TYPE ggml-${backend} TYPE) get_target_property(CURRENT_BACKEND_TYPE ggml-${backend} TYPE)
if (CURRENT_BACKEND_TYPE STREQUAL "MODULE_LIBRARY") if (CURRENT_BACKEND_TYPE STREQUAL "MODULE_LIBRARY")
@@ -90,6 +92,11 @@ macro(link_ggml_backends target)
target_link_libraries(${target} PRIVATE ggml-${backend}) target_link_libraries(${target} PRIVATE ggml-${backend})
endif() endif()
endforeach() endforeach()
# SYCL backend links the runtime PRIVATE, so consumers need -fsycl
# on their own link line to resolve libsycl.so symbols.
if(TARGET ggml-sycl AND GGML_SYCL)
target_link_options(${target} PRIVATE -fsycl)
endif()
add_dependencies(${target} version) add_dependencies(${target} version)
endmacro() endmacro()
buildsycl.sh
Executable
+8
View File
@@ -0,0 +1,8 @@
#!/bin/bash
rm -rf build
mkdir build
cd build
cmake .. -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
cmake --build . --config Release -j "$(nproc)"
tests/tts-SYCL0-Q8_0.log
+90
View File
@@ -0,0 +1,90 @@
# Intel Arc A310 (DG2, 4 GB VRAM) — SYCL0 backend
# oneAPI 2026.0, Level Zero 1.6.33578+15, Ubuntu 24.04 LXC
# Model: qwen-talker-0.6b-base-Q8_0.gguf + qwen-tokenizer-12hz-Q8_0.gguf
# Input: "Hello world. This is a test of Intel Arc GPU acceleration for text to speech."
# Seed: 42, max-new: 128, lang: English
[Qwen] qwentts.cpp 55b7234 (2026-05-17)
Build with Macros:
GGML_SYCL_FORCE_MMQ: no
GGML_SYCL_F16: no
GGML_SYCL_GRAPH: yes
GGML_SYCL_DNNL: yes
Running with Environment Variables:
GGML_SYCL_DEBUG: 0
GGML_SYCL_DISABLE_OPT: 0
GGML_SYCL_DISABLE_GRAPH: 1
GGML_SYCL_DISABLE_DNN: DNN disabled by compile flag
GGML_SYCL_PRIORITIZE_DMMV: 0
GGML_SYCL_ENABLE_FLASH_ATTN: 1
Found 1 SYCL devices:
| | | | |Max | |Max |Global | |
| | | | |compute|Max work|sub |mem | |
|ID| Device Type| Name|Version|units |group |group|size | Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]| Intel Arc A310 LP Graphics| 12.56| 96| 1024| 32| 4032M| 1.6.33578+15|
SYCL Optimization Feature:
|ID| Device Type|Reorder|
|--|-------------------|-------|
| 0| [level_zero:gpu:0]| Y|
[Load] Talker backend: SYCL0 (CPU threads: 2)
[GGUF] models/qwen-talker-0.6b-base-Q8_0.gguf: 478 tensors, data at offset 5954368
[WeightCtx] Loaded 316 tensors, 774.6 MB into backend
[Talker] Loaded: 28 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, mrope sections [24,20,20] interleaved=0
[WeightCtx] Loaded 86 tensors, 143.5 MB into backend
[CodePredictor] Loaded: 5 layers, hidden 1024, heads 16/8, head_dim 128, FFN 3072, RoPE theta 1000000, 15 acoustic codebooks (vocab 2048 each), mtp_proj identity
[WeightCtx] Loaded 76 tensors, 22.9 MB into backend
[SpeakerEncoder] Loaded: enc_dim=1024 sr=24000 mel_dim=128 hidden=512 mfa=1536 asp_attn=128 se=128 scale=8
[GGUF] models/qwen-tokenizer-12hz-Q8_0.gguf: 398 tensors, data at offset 29760
[WeightCtx] Loaded 18 tensors, 33.0 MB into backend
[Quantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512
[WeightCtx] Loaded 93 tensors, 30.9 MB into backend
[Transformer] Loaded: 8 layers, hidden 512, heads 16/16, head_dim 64, FFN 1024, RoPE theta 10000, sliding window 72
[WeightCtx] Loaded 22 tensors, 33.1 MB into backend
[Upsample] Loaded: 2 blocks (2x ratio per block), channels 1024, dwconv kernel 7
[WeightCtx] Loaded 118 tensors, 142.9 MB into backend
[DAC] Loaded: 4 blocks (strides 8/5/4/3), 24 kHz mono out, weights 142.9 MB
[WeightCtx] Loaded 2 tensors, 3.0 MB into backend
[WeightCtx] Loaded 28 tensors, 24.1 MB into backend
[SEANet] Loaded: 4 stages (ratios 4/5/6/8), 64 -> 512 channels, weights 24.1 MB
[WeightCtx] Loaded 96 tensors, 25.6 MB into backend
[EncTransformer] Loaded: 8 layers, hidden 512, heads 8/8, head_dim 64, FFN 2048, RoPE theta 10000
[WeightCtx] Loaded 1 tensors, 2.0 MB into backend
[EncDownsample] Loaded: k=4 stride=2, 512 -> 512 channels, weights 2.0 MB
[WeightCtx] Loaded 20 tensors, 34.0 MB into backend
[EncQuantizer] Loaded: 16 codebooks (1 semantic + 15 acoustic), 2048 entries x 256 dim, hidden 512, weights 34.0 MB
[Pipeline] Ready: hop 1920 samples @ 24000 Hz mono, 16 codebooks @ 12.5 Hz
[KVCache] Allocated: 28 layers, 8 KV heads, head_dim 128, max_seq_len 4096 -> 896 MB
[KVCache] Allocated: 5 layers, 8 KV heads, head_dim 128, max_seq_len 16 -> 0 MB
[Pipeline] Loaded: arch=0b6 variant=base tokenizer=qwen3_tts_tokenizer_12hz codebooks=16 speaker_encoder=loaded speakers=0 fa=on clamp_fp16=off
[BPE] Loaded from GGUF: 151676 vocab, 151291 merges, eos_id=151643
[BPE] Registered 5 arch special tokens (total specials=6)
[Prompt] Built: 25 ids, N_text=17, N_instruct=0, T_ctx=27, hidden=1024, lang=English (id=2050), speaker=none (id=-1) ref_spk_emb=no icl=no
[Sample] step=0 c0=1221 u=0.6129598618 subseq=0
[Sample-CP] g=0 c=16 u=0.0100588445 subseq=1
[Sample-CP] g=1 c=580 u=0.3984136879 subseq=2
[Sample-CP] g=2 c=159 u=0.0403083973 subseq=3
[Sample-CP] g=3 c=445 u=0.1562667489 subseq=4
[Sample-CP] g=4 c=651 u=0.4824730754 subseq=5
[Sample-CP] g=5 c=1186 u=0.7362473011 subseq=6
[Sample-CP] g=6 c=886 u=0.4059819579 subseq=7
[Sample-CP] g=7 c=1385 u=0.5188996196 subseq=8
[Sample-CP] g=8 c=224 u=0.2867110968 subseq=9
[Sample-CP] g=9 c=448 u=0.2415804267 subseq=10
[Sample-CP] g=10 c=1930 u=0.9228392243 subseq=11
[Sample-CP] g=11 c=1527 u=0.8299142122 subseq=12
[Sample-CP] g=12 c=32 u=0.0341879092 subseq=13
[Sample-CP] g=13 c=563 u=0.3878940344 subseq=14
[Sample-CP] g=14 c=163 u=0.0823899582 subseq=15
[Sample] step=1 c0=1204 u=0.7741796374 subseq=16
[Pipeline] Generated 8 frames
[Pipeline] Generated 16 frames
[Pipeline] Generated 24 frames
[Pipeline] Generated 32 frames
[Pipeline] Generated 40 frames
[Pipeline] Generated 48 frames
[Pipeline] Generated 56 frames
[Pipeline] EOS at step 63, stopping
[Pipeline] Generation done : 63 frames
[WAV] Wrote /tmp/test-arc.wav: 120960 samples, 24000 Hz, mono S16
[Pipeline] Wrote 120960 samples (5.04 s) -> /tmp/test-arc.wav