Chi_Liu AmosLewis

MLIR | Deep Learning | Computer Vision

AmosLewis / bisect-log-3.12-3.13-00e-445.sh

Created March 17, 2025 01:17

	# downdload the issue1.sh by the following wget
	# wget https://gist.githubusercontent.com/AmosLewis/00fdb4e9a96f29c188828e3ff4ea29ef/raw/8377ffd2f53e58e12a9adae7b92b3d5a7f35d98b/bisect-issue1.sh


	(bisect.venv) ➜ bisect git:(main) ✗ python ./bisect_packages.py \
	--good-ref=00e88733e6b8c8cdb351d4516509f56daebdf604 \
	--bad-ref=4451b8ba42b1249eb3ed1b5031a46f51733984ed \
	--test-script=/sharedfile/attn/bisect/issue1.sh
	Welcome to bisect_packages.py!

AmosLewis / bisect-issue1.sh

Last active March 17, 2025 01:06

	# wget https://sharkpublic.blob.core.windows.net/sharkpublic/chi/llama/atten/fp8_attn.mlir
	iree-compile \
	/sharedfile/attn/128/fp8_attn.mlir \
	--iree-hip-target=gfx942 \
	-o=/sharedfile/attn/128/fp8_attn.vmfb \
	--iree-hal-target-device=hip \
	--iree-dispatch-creation-enable-aggressive-fusion=true \
	--iree-global-opt-propagate-transposes=true \
	--iree-opt-aggressively-propagate-transposes=true \
	--iree-opt-data-tiling=false \

AmosLewis / bisect-log-3.11-3.12-6ea-00e.log

Last active March 17, 2025 01:01

	# downdload the issue1.sh by the following wget
	# wget https://gist.githubusercontent.com/AmosLewis/00fdb4e9a96f29c188828e3ff4ea29ef/raw/8377ffd2f53e58e12a9adae7b92b3d5a7f35d98b/bisect-issue1.sh

	(bisect.venv) ➜ bisect git:(main) ✗ python ./bisect_packages.py \
	--good-ref=6eadf3df6f57a20283da8318f9178b0d92ae50eb \
	--bad-ref=00e88733e6b8c8cdb351d4516509f56daebdf604 \
	--test-script=/sharedfile/attn/bisect/issue1.sh
	Welcome to bisect_packages.py!

	------------------------------------------------------------------

AmosLewis / 20106+block_VS_before20106+block.log

Last active March 4, 2025 17:35

	/home/chi/src/iree-build-trace/tools/iree-compile \
	/sharedfile/attn/128/fp8_attn.mlir \
	--iree-hip-target=gfx942 \
	-o=/sharedfile/attn/128/fp8_attn.vmfb \
	--iree-hal-target-device=hip \
	--iree-dispatch-creation-enable-aggressive-fusion=true \
	--iree-global-opt-propagate-transposes=true \
	--iree-opt-aggressively-propagate-transposes=true \
	--iree-opt-data-tiling=false \
	--iree-preprocessing-pass-pipeline='builtin.module(util.func(iree-preprocessing-generalize-linalg-matmul-experimental))' \

AmosLewis / attn8_mask_export_bug.txt

Last active February 26, 2025 18:26

	(.venv) ➜ 32 python3 -m sharktank.examples.export_paged_llm_v1 --irpa-file=/sharedfile/llama3_8b_fp8.irpa \
	--output-mlir=/sharedfile/32/fp8_attn.mlir \
	--output-config=/sharedfile/32/config_attn.json \
	--bs=1 --attention-kernel sharktank \
	--attention-dtype=bfloat16 \
	--activation-dtype=bfloat16 \
	--kv-cache-dtype=float8_e4m3fnuz \
	--use-hf \
	--use-attention-mask

AmosLewis / run_bm_0224mlir_ir0214_0215_0224.log

Last active February 24, 2025 22:40

	iree-base-compiler 3.3.0rc20250215
	iree-base-runtime 3.3.0rc20250215
	iree-turbine 3.3.0rc20250215

	iree-compile /sharedfile/128/fp8_128.mlir \
	--iree-hip-target=gfx942 \
	-o=/sharedfile/128/fp8_128_0224_ir0215.vmfb \
	--iree-hal-target-device=hip \
	--iree-dispatch-creation-enable-aggressive-fusion=true \
	--iree-global-opt-propagate-transposes=true \

AmosLewis / run_bm_0212mlir_ir0209_ir0224.log

Last active February 24, 2025 22:40

	iree-base-compiler 3.2.0rc20250209
	iree-base-runtime 3.2.0rc20250209

	iree-compile /sharedfile/128/fp8_128_0212.mlir \
	--iree-hip-target=gfx942 \
	-o=/sharedfile/128/fp8_128_0212_ir0209.vmfb \
	--iree-hal-target-device=hip \
	--iree-dispatch-creation-enable-aggressive-fusion=true \
	--iree-global-opt-propagate-transposes=true \
	--iree-opt-aggressively-propagate-transposes=true \

AmosLewis / llama_fp8_compile_bug.txt

Last active February 24, 2025 18:04

kv16

This file has been truncated, but you can view the full file.

	iree-base-compiler 3.3.0rc20250223
	iree-base-runtime 3.3.0rc20250223
	iree-turbine 3.3.0rc20250223

	wget https://sharkpublic.blob.core.windows.net/sharkpublic/chi/llama/fp8_32_kv16.mlir

	iree-compile /sharedfile/32/fp8_32_kv16.mlir \
	--iree-hip-target=gfx942 \
	-o=/sharedfile/32/fp8_32_kv16.vmfb \
	--iree-hal-target-device=hip \

AmosLewis / llama_fp8_export_bug.txt

Created February 24, 2025 17:14

	python3 -m sharktank.examples.export_paged_llm_v1 --irpa-file=/sharedfile/llama3_8b_fp8.irpa \
	--output-mlir=/sharedfile/32/fp8_32.mlir \
	--output-config=/sharedfile/32/config_32.json \
	--bs=1 --attention-kernel torch \
	--attention-dtype=float8_e4m3fnuz --activation-dtype=bfloat16 \
	--use-hf \
	--kv-cache-dtype=float8_e4m3fnuz
	/home/chi/src/shark-ai/.venv/lib/python3.11/site-packages/iree/turbine/aot/params.py:163: UserWarning: The given NumPy array is not writable, and PyTorch does not support non-writable tensors. This means writing to this tensor will result in undefined behavior. You may want to copy the array to protect its data or make it writable before converting it to a tensor. This type of warning will be suppressed for the rest of this program. (Triggered internally at ../torch/csrc/utils/tensor_numpy.cpp:206.)
	return torch.from_numpy(wrapper)
	Exporting prefill_bs1

AmosLewis / llama_fp8_attn8_bs4_input128.mlir

Created February 19, 2025 20:40

This file has been truncated, but you can view the full file.