ggml : support `bs > 512` for Metal `ggml_mul_mat_id`

[stewartoallen]

Mixtral models + metal gpu + batch size > 512 = GGML_ASERT. Does not affect models such as llama-2-7b-chat.Q5_K_M.gguf

Hardware: Apple M2 Ultra
RAM: 192GB
llama.cpp current version as of 2024-01-21 (504dc37)

./main -f /tmp/prompt1k -m models/mixtral-8x7b-instruct-v0.1.Q6_K.gguf -c 4096 -b 512 << OK
./main -f /tmp/prompt1k -m models/mixtral-8x7b-instruct-v0.1.Q6_K.gguf -c 4096 -b 4096 << FAIL

### Assistant:GGML_ASSERT: ggml-metal.m:1511: ne11 <= 512

./main -f /tmp/prompt1k -m models/mixtral-8x7b-instruct-v0.1.Q6_K.gguf -c 4096 -b 4096 -ngl 0 << OK

but takes forever

[ggerganov]

Yes, need to move the src1ids from the stack to shared memory and increase it's size

[xcottos]

Hi everybody,

I'm encountering the same using the python wrapper (the python kernel crashes if batch_size > 512):

GGML_ASSERT: /private/var/folders/md/5gb2vml53fl36jdz9tvg53s80000gn/T/pip-install-b5i4cgto/llama-cpp-python_1549dcff18604e30944aeaa6c55a63b3/vendor/llama.cpp/ggml-metal.m:1726: ne11 <= 512
GGML_ASSERT: /private/var/folders/md/5gb2vml53fl36jdz9tvg53s80000gn/T/pip-install-b5i4cgto/llama-cpp-python_1549dcff18604e30944aeaa6c55a63b3/vendor/llama.cpp/ggml-metal.m:1726: ne11 <= 512
GGML_ASSERT: /private/var/folders/md/5gb2vml53fl36jdz9tvg53s80000gn/T/pip-install-b5i4cgto/llama-cpp-python_1549dcff18604e30944aeaa6c55a63b3/vendor/llama.cpp/ggml-metal.m:1726: ne11 <= 512
GGML_ASSERT: /private/var/folders/md/5gb2vml53fl36jdz9tvg53s80000gn/T/pip-install-b5i4cgto/llama-cpp-python_1549dcff18604e30944aeaa6c55a63b3/vendor/llama.cpp/ggml-metal.m:1726: ne11 <= 512
zsh: abort /Volumes/AI_MASTER/envs/torch-metal/bin/python
/Volumes/AI_MASTER/envs/torch-metal/lib/python3.11/multiprocessing/resource_tracker.py:224: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
warnings.warn('resource_tracker: There appear to be %d '

I can use the batch_size max to 512, but I'm sure it was working before, so not sure what update broke it...

Please advise if I'm missing any update or procedure to address it

Thanks
Luca

[xcottos]

I just verified that in the llama.cpp version I have (at least 3 weeks behind) this works:

./main -i -m /Volumes/AI_MASTER/models/mistral-8x7b-instruct/mixtral-8x7b-instruct-v0.1.Q4_K_M.gguf -c 8196 -ngl 1 -b 8196 -t 0 --color -p "[INST] what is the capital of france? [/INST]"

..............
ggml_metal_init: recommendedMaxWorkingSetSize = 51539.61 MB
ggml_metal_init: maxTransferRate = built-in GPU
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size = 1024.50 MiB, (26241.81 / 49152.00)
llama_new_context_with_model: KV self size = 1024.50 MiB, K (f16): 512.25 MiB, V (f16): 512.25 MiB
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size = 0.02 MiB, (26241.83 / 49152.00)
llama_build_graph: non-view tensors processed: 1124/1124
llama_new_context_with_model: compute buffer total size = 9292.41 MiB
ggml_backend_metal_buffer_type_alloc_buffer: allocated buffer, size = 9289.23 MiB, (35531.05 / 49152.00)

system_info: n_threads = 10 / 10 | AVX = 0 | AVX2 = 0 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | FMA = 0 | NEON = 1 | ARM_FMA = 1 | F16C = 0 | FP16_VA = 1 | WASM_SIMD = 0 | BLAS = 1 | SSE3 = 0 | SSSE3 = 0 | VSX = 0 |
main: interactive mode on.
sampling:
repeat_last_n = 64, repeat_penalty = 1.100, frequency_penalty = 0.000, presence_penalty = 0.000
top_k = 40, tfs_z = 1.000, top_p = 0.950, min_p = 0.050, typical_p = 1.000, temp = 0.800
mirostat = 0, mirostat_lr = 0.100, mirostat_ent = 5.000
sampling order:
CFG -> Penalties -> top_k -> tfs_z -> typical_p -> top_p -> min_p -> temp
generate: n_ctx = 8196, n_batch = 8196, n_predict = -1, n_keep = 0

== Running in interactive mode. ==

• Press Ctrl+C to interject at any time.
• Press Return to return control to LLaMa.
• To return control without starting a new line, end your input with '/'.
• If you want to submit another line, end your input with ''.

[INST] what is the capital of france? [/INST] The capital city of France is Paris. It is located in the northern central part of the country, on the river Seine. Paris is known for its iconic landmarks such as the Eiffel Tower, Notre-Dame Cathedral, and the Louvre Museum, which is home to thousands of works of art, including the Mona Lisa. Paris is also famous for its fashion, cuisine, and cultural events, making it one of the most popular tourist destinations in the world.

Now I'm scared of updating llama.cpp since recompiling llama-python-cpp ended up in the behaviour described above.

please advise how I can address it

Thank you
Luca

[ggerganov]

Use n_batch <= 512 for now - this change was made to improve the performance. Larger batches will be supported in the future (#5070 (comment))

[xcottos]

Thank you for the update

Luca

[weissenbacherpwc]

same issue here. @ggerganov looking forward to the changes in the future