inconsistent dev-requirements.txt when pyproject.toml ? xprof==2.21.3

[truatpasteurdotfr]

hello

I am trying to rebuild alphafold3 on dgx-spark according to #394 but I building fails with

Collecting wheel==0.46.1 (from -r dev-requirements.txt (line 2313))
  Downloading wheel-0.46.1-py3-none-any.whl (23 kB)
Collecting wrapt==2.0.1 (from -r dev-requirements.txt (line 2317))
  Downloading wrapt-2.0.1-cp312-cp312-manylinux2014_aarch64.manylinux_2_17_aarch64.manylinux_2_28_aarch64.whl (123 kB)
ERROR: Ignored the following yanked versions: 2.20.3
ERROR: Ignored the following versions that require a different python version: 1.10.0 Requires-Python >=3.8,<3.12; 1.10.0rc1 Requires-Python >=3.8,<3.12; 1.10.0rc2 Requires-Python >=3.8,<3.12; 1.10.1 Requires-Python >=3.8,<3.12; 1.21.2 Requires-Python >=3.7,<3.11; 1.21.3 Requires-Python >=3.7,<3.11; 1.21.4 Requires-Python >=3.7,<3.11; 1.21.5 Requires-Python >=3.7,<3.11; 1.21.6 Requires-Python >=3.7,<3.11; 1.6.2 Requires-Python >=3.7,<3.10; 1.6.3 Requires-Python >=3.7,<3.10; 1.7.0 Requires-Python >=3.7,<3.10; 1.7.1 Requires-Python >=3.7,<3.10; 1.7.2 Requires-Python >=3.7,<3.11; 1.7.3 Requires-Python >=3.7,<3.11; 1.8.0 Requires-Python >=3.8,<3.11; 1.8.0rc1 Requires-Python >=3.8,<3.11; 1.8.0rc2 Requires-Python >=3.8,<3.11; 1.8.0rc3 Requires-Python >=3.8,<3.11; 1.8.0rc4 Requires-Python >=3.8,<3.11; 1.8.1 Requires-Python >=3.8,<3.11; 1.9.0 Requires-Python >=3.8,<3.12; 1.9.0rc1 Requires-Python >=3.8,<3.12; 1.9.0rc2 Requires-Python >=3.8,<3.12; 1.9.0rc3 Requires-Python >=3.8,<3.12; 1.9.1 Requires-Python >=3.8,<3.12
ERROR: Could not find a version that satisfies the requirement xprof==2.21.3 (from versions: 2.19.6, 2.19.7, 2.19.8, 2.19.9, 2.20.0, 2.20.2a0, 2.20.2, 2.20.4, 2.20.5, 2.20.6, 2.20.7)
ERROR: No matching distribution found for xprof==2.21.3

I have regenerated the dev-requirements.txt from pyproject.toml and it differs from the one on github.

wget https://raw.githubusercontent.com/google-deepmind/alphafold3/refs/heads/main/pyproject.toml
# sha256sum
# 1bdd1c29573851367de096bce343c19d83047cf65ac4732ba12cbc695d8354ba  pyproject.toml

wget https://raw.githubusercontent.com/google-deepmind/alphafold3/refs/heads/main/dev-requirements.txt -O dev-requirements.txt.github
# sha256sum
aa9b5ea2fa53e4afea7dce46fbc801e241f43d467651b11cc10013e9e6401493  dev-requirements.txt.github

uv pip compile --generate-hashes -o dev-requirements.txt pyproject.toml

From my limited knowledge, the main difference is xprof==2.20.7 VS xprof==2.21.3

dev-requirements.txt

With the version attached I could build af3.

[Augustin-Zidek]

Hello, you seem to be using Python 3.11 on your system (based on the generated dev-requirements.txt). Could you try with Python 3.12?

[truatpasteurdotfr]

I was using the python3 from nvidia/cuda:12.6.3-base-ubuntu24.04, which is 3.12.3-0ubuntu2.1.

[Augustin-Zidek]

Hmm, let me fix this properly by migrating to modern uv-based setup. Are you ok using the workaround with the fixed xprof version?

[truatpasteurdotfr]

yes, it builds fine on the DGX-spark for me, but seems quite slow for 2PV7

real    17m24.212s
user    37m47.040s
sys     4m45.819s

compared to #434 (comment)

[Augustin-Zidek]

@truatpasteurdotfr is this including the MSA/template run time, or inference only? If it was including MSA/template search (i.e. data pipeline + inference), than the runtime looks normal to me.

[truatpasteurdotfr]

yes that incluse the MSA search (on local nvme storage) and inference time, thanks for your feedback!

[Augustin-Zidek]

Could you try updating past 8d37fc1? I modernized the setup and migrated to uv.

[truatpasteurdotfr]

Here is the build log attached from:

log.txt

sudo docker build -t af3 -f docker/Dockerfile  . 2>&1 | tee log.txt

[Augustin-Zidek]

Ah, this is because you are on an arm64 machine. Could you try doing what is suggested in the error message:

You're on Linux (`manylinux_2_39_aarch64`), but `xprof` (v2.21.3) only has wheels
for the following platforms: `manylinux2014_x86_64`, `macosx_12_0_arm64`;
consider adding "sys_platform == 'linux' and platform_machine == 'aarch64'" to 
`tool.uv.required-environments` to ensure uv resolves to a version with compatible wheels

Also, does changing the line in Dockerfile from:

uv sync --frozen --all-groups --no-editable

to

uv sync --all-groups --no-editable

fix the issue?

[truatpasteurdotfr]

right the DGX-spark is ARM64, adding the requirement and changing the uv sync... line works:

tru@spark-50df:~/git/github/google-deepmind/alphafold3$ git diff
diff --git a/docker/Dockerfile b/docker/Dockerfile
index 7b413af..a71174e 100644
--- a/docker/Dockerfile
+++ b/docker/Dockerfile
@@ -70,7 +70,7 @@ WORKDIR /app/alphafold
 # AlphaFold 3 code and change its dependencies, remove the --frozen and
 # --no-editable flags.
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv sync --frozen --all-groups --no-editable
+    uv sync --all-groups --no-editable
 
 # Build chemical components database (this binary was installed by uv sync).
 RUN uv run build_data
diff --git a/pyproject.toml b/pyproject.toml
index 614e05b..9735e33 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -34,6 +34,9 @@ dev = [
 
 [tool.uv]
 package = true
+required-environments = [
+    "sys_platform == 'linux' and platform_machine == 'aarch64'"
+]
 
 [tool.scikit-build]
 wheel.exclude = [

tru@spark-50df:~/git/github/google-deepmind/alphafold3$ sudo docker build -t af3 -f docker/Dockerfile  . 2>&1 | tee log2.txt

attached log2.txt file

log2.txt

[Augustin-Zidek]

Thanks. Were both of these changes needed, or was the first one sufficient?

[Augustin-Zidek]

This should fix the problem: 22b9ab8. Could you test it with your system?

[Landau1994]

This should fix the problem: 22b9ab8. Could you test it with your system?

Thanks to the authors kindly support~~~~
I have encountered the same issue, 22b9ab8, works. And I have also figured out a slightly different installation strategy for dgx-spark:

git clone https://github.com/google-deepmind/alphafold3.git

conda create -n af3_new python=3.12
pip install uv

### compile from source as AF3 official guide and compile the jackhmmer_seq_limit.patch from the official code

bash hmmer_install.sh


uv pip install xprof
## it's verion will be 2.20.7
## In my opion the version of xprof is not a big issue for installing and using successfully
## delete xprof line in uv.lock to avoid uv sync bug

uv sync --frozen --all-groups --no-editable
uv run build_data

## test install works
uv run python run_alphafold.py

The hmmer_install.sh is adapted from AF3 official docker file

#!/bin/bash

BUILD=$HOME/build/hmmer_build
PREFIX=$HOME/software/hmmer
PATCH=$HOME/software/AF3/alphafold3/docker/jackhmmer_seq_limit.patch

mkdir -p "$BUILD" "$PREFIX"
cd "$BUILD"

wget http://eddylab.org/software/hmmer/hmmer-3.4.tar.gz
echo "ca70d94fd0cf271bd7063423aabb116d42de533117343a9b27a65c17ff06fbf3  hmmer-3.4.tar.gz" | sha256sum --check

tar zxf hmmer-3.4.tar.gz
rm hmmer-3.4.tar.gz

patch -p0 < "$PATCH"

cd hmmer-3.4
./configure --prefix=$CONDA_PREFIX

make -j$(nproc)
make install
cd easel && make install

#rm -R $BUILD

Now with uv to help manage the environment, if you want to run AF3 in a different directory, here is my way:

#!/bin/bash

#export TRITON_PTXAS_PATH=/usr/local/cuda-13.0/bin/ptxas
#export CUDA_HOME=/usr/local/cuda-13.0
#export PATH=/usr/local/cuda-13.0/bin:$PATH
#export LD_LIBRARY_PATH=/usr/local/cuda-13.0/lib64:${LD_LIBRARY_PATH}

export XLA_FLAGS="--xla_gpu_enable_triton_gemm=false"
# I test and configure this the optimal configure for dgx-spark, preallocate will be slow
#export XLA_PYTHON_CLIENT_PREALLOCATE=true
#export XLA_CLIENT_MEM_FRACTION=0.95

AF3_DIR=~/software/AF3/alphafold3
MODEL=~/software/AF3/weights/
DB=~/software/AF3/database/

PROJECT_DIR=~/software/AF3/demo_20260121
INPUT=$PROJECT_DIR/input/2PV7_data_raw.json
OUTPUT=$PROJECT_DIR/output

JAX_ENABLE_X64=1

uv run --project $AF3_DIR \
       	python $AF3_DIR/run_alphafold.py \
	--json_path=$INPUT \
	--model_dir=$MODEL \
	--db_dir=$DB \
	--output_dir=$OUTPUT

As I have tested before, if you have already performed alignment, it will take 4 minutes for the 2PV7 demo.

In this case, I tested with sequence alignment, which takes around 10 minutes. Here are the complete logs

/home/wlt2025/software/AF3/alphafold3/.venv/lib/python3.12/site-packages/keras/src/export/tf2onnx_lib.py:8: FutureWarning: In the future `np.object` will be defined as the corresponding NumPy scalar.
  if not hasattr(np, "object"):
INFO:2026-01-21 02:32:19,529:jax._src.xla_bridge:812: Unable to initialize backend 'tpu': INTERNAL: Failed to open libtpu.so: libtpu.so: cannot open shared object file: No such file or directory
I0121 02:32:19.529305 260505248459360 xla_bridge.py:812] Unable to initialize backend 'tpu': INTERNAL: Failed to open libtpu.so: libtpu.so: cannot open shared object file: No such file or directory
W0121 02:32:19.722134  110496 cuda_executor.cc:1802] GPU interconnect information not available: INTERNAL: NVML doesn't support extracting fabric info or NVLink is not used by the device.
W0121 02:32:19.726824  110413 cuda_executor.cc:1802] GPU interconnect information not available: INTERNAL: NVML doesn't support extracting fabric info or NVLink is not used by the device.
I0121 02:32:25.422684 260505248459360 pipeline.py:82] Getting protein MSAs for sequence GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG
I0121 02:32:25.424953 260485790167424 jackhmmer.py:177] Query sequence (non-sharded db): GMRESYANENQFGFKT... (len 298)
I0121 02:32:25.425163 260485773259136 jackhmmer.py:177] Query sequence (non-sharded db): GMRESYANENQFGFKT... (len 298)
I0121 02:32:25.425472 260485781713280 jackhmmer.py:177] Query sequence (non-sharded db): GMRESYANENQFGFKT... (len 298)
I0121 02:32:25.425740 260485798621568 jackhmmer.py:177] Query sequence (non-sharded db): GMRESYANENQFGFKT... (len 298)
I0121 02:32:25.425840 260485790167424 subprocess_utils.py:82] Launching subprocess "/home/wlt2025/anaconda3/envs/af3_new/bin/jackhmmer -o /dev/null -A /tmp/tmp1k259uyv/output.sto --noali --F1 0.0005 --F2 5e-05 --F3 5e-07 --cpu 8 -N 1 -E 0.0001 --incE 0.0001 --seq_limit 5000 /tmp/tmp1k259uyv/query.fasta /home/wlt2025/software/AF3/database//mgy_clusters_2022_05.fa"
I0121 02:32:25.425903 260485773259136 subprocess_utils.py:82] Launching subprocess "/home/wlt2025/anaconda3/envs/af3_new/bin/jackhmmer -o /dev/null -A /tmp/tmpyz9cn1l7/output.sto --noali --F1 0.0005 --F2 5e-05 --F3 5e-07 --cpu 8 -N 1 -E 0.0001 --incE 0.0001 --seq_limit 50000 /tmp/tmpyz9cn1l7/query.fasta /home/wlt2025/software/AF3/database//uniprot_all_2021_04.fa"
I0121 02:32:25.426295 260485781713280 subprocess_utils.py:82] Launching subprocess "/home/wlt2025/anaconda3/envs/af3_new/bin/jackhmmer -o /dev/null -A /tmp/tmpi6rb88ho/output.sto --noali --F1 0.0005 --F2 5e-05 --F3 5e-07 --cpu 8 -N 1 -E 0.0001 --incE 0.0001 --seq_limit 5000 /tmp/tmpi6rb88ho/query.fasta /home/wlt2025/software/AF3/database//bfd-first_non_consensus_sequences.fasta"
I0121 02:32:25.426715 260485798621568 subprocess_utils.py:82] Launching subprocess "/home/wlt2025/anaconda3/envs/af3_new/bin/jackhmmer -o /dev/null -A /tmp/tmpdzl16bcx/output.sto --noali --F1 0.0005 --F2 5e-05 --F3 5e-07 --cpu 8 -N 1 -E 0.0001 --incE 0.0001 --seq_limit 10000 /tmp/tmpdzl16bcx/query.fasta /home/wlt2025/software/AF3/database//uniref90_2022_05.fa"
I0121 02:33:09.457022 260485781713280 subprocess_utils.py:111] Finished Jackhmmer (bfd-first_non_consensus_sequences.fasta) in 44.031 seconds
I0121 02:35:09.129067 260485798621568 subprocess_utils.py:111] Finished Jackhmmer (uniref90_2022_05.fa) in 163.702 seconds
I0121 02:36:33.951113 260485773259136 subprocess_utils.py:111] Finished Jackhmmer (uniprot_all_2021_04.fa) in 248.525 seconds
I0121 02:37:39.313379 260485790167424 subprocess_utils.py:111] Finished Jackhmmer (mgy_clusters_2022_05.fa) in 313.887 seconds
I0121 02:37:39.393098 260505248459360 pipeline.py:115] Getting protein MSAs took 313.97 seconds for sequence GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG
I0121 02:37:39.393228 260505248459360 pipeline.py:121] Deduplicating MSAs for sequence GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG
I0121 02:37:39.411430 260505248459360 pipeline.py:134] Deduplicating MSAs took 0.02 seconds for sequence GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG, found 8506 unpaired sequences, 7080 paired sequences
I0121 02:37:39.414676 260505248459360 pipeline.py:40] Getting protein templates for sequence GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG
I0121 02:37:39.470546 260505248459360 subprocess_utils.py:82] Launching subprocess "/home/wlt2025/anaconda3/envs/af3_new/bin/hmmbuild --informat stockholm --hand --amino /tmp/tmpsf_gh8tb/output.hmm /tmp/tmpsf_gh8tb/query.msa"
I0121 02:37:39.676601 260505248459360 subprocess_utils.py:111] Finished Hmmbuild in 0.206 seconds
I0121 02:37:39.679155 260505248459360 subprocess_utils.py:82] Launching subprocess "/home/wlt2025/anaconda3/envs/af3_new/bin/hmmsearch --noali --cpu 8 --F1 0.1 --F2 0.1 --F3 0.1 -E 100 --incE 100 --domE 100 --incdomE 100 -A /tmp/tmpdohu07mc/output.sto /tmp/tmpdohu07mc/query.hmm /home/wlt2025/software/AF3/database/pdb_seqres_2022_09_28.fasta"
I0121 02:37:44.627124 260505248459360 subprocess_utils.py:111] Finished Hmmsearch (pdb_seqres_2022_09_28.fasta) in 4.948 seconds
I0121 02:37:44.835514 260505248459360 pipeline.py:52] Getting 4 protein templates took 5.42 seconds for sequence GMRESYANENQFGFKTINSDIHKIVIVGGYGKLGGLFARYLRASGYPISILDREDWAVAESILANADVVIVSVPINLTLETIERLKPYLTENMLLADLTSVKREPLAKMLEVHTGAVLGLHPMFGADIASMAKQVVVRCDGRFPERYEWLLEQIQIWGAKIYQTNATEHDHNMTYIQALRHFSTFANGLHLSKQPINLANLLALSSPIYRLELAMIGRLFAQDAELYADIIMDKSENLAVIETLKQTYDEALTFFENNDRQGFIDAFHKVRDWFGDYSEQFLKESRQLLQQANDLKQG
I0121 02:37:47.894139 260505248459360 pipeline.py:173] processing 2PV7, random_seed=1
I0121 02:37:48.201773 260505248459360 pipeline.py:266] Calculating bucket size for input with 596 tokens.
I0121 02:37:48.202036 260505248459360 pipeline.py:272] Got bucket size 768 for input with 596 tokens, resulting in 172 padded tokens.
I0121 02:37:52.245805 260505248459360 cache.py:95] Autotuning cache file not found: /home/wlt2025/software/AF3/alphafold3/.venv/lib/python3.12/site-packages/tokamax/data/autotuning/nvidia_gb10/pallas_triton_gated_linear_unit.json
I0121 02:37:52.339859 260505248459360 cache.py:95] Autotuning cache file not found: /home/wlt2025/software/AF3/alphafold3/.venv/lib/python3.12/site-packages/tokamax/data/autotuning/nvidia_gb10/pallas_triton_flash_attention.json
/home/wlt2025/software/AF3/alphafold3/src/alphafold3/model/network/featurization.py:104: UserWarning: Explicitly requested dtype int64 requested in broadcasted_iota is not available, and will be truncated to dtype int32. To enable more dtypes, set the jax_enable_x64 configuration option or the JAX_ENABLE_X64 shell environment variable. See https://github.com/jax-ml/jax#current-gotchas for more.
  iota = jax.lax.broadcasted_iota(jnp.int64, logits.shape, axis)

Running AlphaFold 3. Please note that standard AlphaFold 3 model parameters are
only available under terms of use provided at
https://github.com/google-deepmind/alphafold3/blob/main/WEIGHTS_TERMS_OF_USE.md.
If you do not agree to these terms and are using AlphaFold 3 derived model
parameters, cancel execution of AlphaFold 3 inference with CTRL-C, and do not
use the model parameters.

Found local devices: [CudaDevice(id=0)], using device 0: cuda:0
Building model from scratch...
Checking that model parameters can be loaded...

Running fold job 2PV7...
Output will be written in /home/wlt2025/software/AF3/demo_20260121/output/2PV7_20260121_023225 since /home/wlt2025/software/AF3/demo_20260121/output/2PV7 is non-empty.
Running data pipeline...
Running data pipeline for chain A...
Running data pipeline for chain A took 319.45 seconds
Running data pipeline for chain B...
Running data pipeline for chain B took 0.03 seconds
Writing model input JSON to /home/wlt2025/software/AF3/demo_20260121/output/2PV7_20260121_023225/2PV7_data.json
Predicting 3D structure for 2PV7 with 1 seed(s)...
Featurising data with 1 seed(s)...
Featurising data with seed 1.
Featurising data with seed 1 took 4.26 seconds.
Featurising data with 1 seed(s) took 7.23 seconds.
Running model inference and extracting output structure samples with 1 seed(s)...
Running model inference with seed 1...
Running model inference with seed 1 took 188.52 seconds.
Extracting inference results with seed 1...
Extracting 5 inference samples with seed 1 took 0.28 seconds.
Running model inference and extracting output structures with 1 seed(s) took 188.80 seconds.
Writing outputs with 1 seed(s)...
Fold job 2PV7 done, output written to /home/wlt2025/software/AF3/demo_20260121/output/2PV7_20260121_023225

Done running 1 fold jobs.

Here is the ranking score of this run:

seed,sample,ranking_score
1,0,0.882906180764509
1,1,0.8909533422906442
1,2,0.8926564700543095
1,3,0.8971755655261637
1,4,0.9036613789172122