XGBoost 3.0.0 Release

[jakirkham]

Recently XGBoost tagged 3.0.0

Filing this to track updating to XGBoost 3.0.0 in conda-forge

xref: dmlc/xgboost#11286

[jakirkham]

Typically the bot will open a version update PR

However that has not happened in this case as the bot runs into the following error on CI

Traceback (most recent call last):
    File "/home/runner/work/cf-scripts/cf-scripts/cf-scripts/conda_forge_tick/auto_tick.py", line 744, in _run_migrator_on_feedstock_branch
      migrator_uid, pr_json = run_with_tmpdir(
                              ^^^^^^^^^^^^^^^^
    File "/home/runner/work/cf-scripts/cf-scripts/cf-scripts/conda_forge_tick/auto_tick.py", line 451, in run_with_tmpdir
      return run(
             ^^^^
    File "/home/runner/work/cf-scripts/cf-scripts/cf-scripts/conda_forge_tick/auto_tick.py", line 533, in run
      migration_run_data = run_migration(
                           ^^^^^^^^^^^^^^
    File "/home/runner/work/cf-scripts/cf-scripts/cf-scripts/conda_forge_tick/migration_runner.py", line 68, in run_migration
      return run_migration_containerized(
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    File "/home/runner/work/cf-scripts/cf-scripts/cf-scripts/conda_forge_tick/migration_runner.py", line 165, in run_migration_containerized
      data = run_container_operation(
             ^^^^^^^^^^^^^^^^^^^^^^^^
    File "/home/runner/micromamba/envs/cf-scripts/lib/python3.11/site-packages/conda_forge_feedstock_ops/container_utils.py", line 264, in run_container_operation
      raise ContainerRuntimeError(
  conda_forge_feedstock_ops.container_utils.ContainerRuntimeError: Error running 'conda-forge-tick-container migrate-feedstock --feedstock-name xgboost --default-branch main --existing-feedstock-node-attrs - --log-level info --kwargs {
    "hash_type": "sha256"
  }' in container - error VersionMigrationError raised:
  "The recipe did not change in the version migration, a URL did not hash, or there is jinja2 syntax the bot cannot handle!
  
  Please check the URLs in your recipe with version '3.0.0' to make sure they exist!
  
  We also found the following errors:
  
   - could not hash URL template '[https://github.com/dmlc/xgboost/releases/download/v{{](https://github.com/dmlc/xgboost/releases/download/v%7B%7B) version }}/xgboost-{{ version }}.tar.gz'
  "

The core of the error is the bot detects 3.0.0 and tries to fill that into the source/url

xgboost-feedstock/recipe/meta.yaml

Lines 10 to 11 in 7517233

	source:
	url: https://github.com/dmlc/xgboost/releases/download/v{{ version }}/xgboost-{{ version }}.tar.gz

However the resulting URL is invalid. This is because the source tarball now has a -src suffix ( dmlc/xgboost#11055 ), which the URL above doesn't include

So we will need to make the update manually and fix the URL at the same time

[trivialfis]

Note: XGB 3.0 pypi build is using CTK12.8, without custom CCCL.

[jakirkham]

Thanks Jiaming! 🙏

During the XGBoost 2.1.0 rollout ( #167 ) we needed to use a newer version of CCCL. So we are using an otherwise unpatched CCCL that matches the latest release on GitHub ( currently 2.8.2: conda-forge/cccl-feedstock#26 ). IIRC this was done to resolve some build issues we saw on CI

xgboost-feedstock/recipe/meta.yaml

Line 46 in 7517233

- cccl !=2.4.0,!=2.5.0 # [cuda_compiler != "None"]

We could try dropping this when upgrading to XGBoost 3

[bdice]

CUDA 12.8 and RAPIDS 25.04 both use CCCL 2.7.0 at the moment, so it should be okay to use CCCL 2.7.0. (RAPIDS will update to 2.8.x soon, but probably for 25.06 and not 25.04.)

[jakirkham]

Sounds good. We were using CCCL 2.7. Though didn't pin it. Have pinned it now to ensure we are aligned (and workaround issues we saw with CCCL 2.8)

xgboost-feedstock/recipe/meta.yaml

Line 51 in f97eb89

- cccl =2.7 # [cuda_compiler != "None"]

[jakirkham]

@conda-forge-admin , please re-render

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-webservice.

I just wanted to let you know that I started rerendering the recipe in #225.

[jakirkham]

@conda-forge-admin , please re-render

[conda-forge-webservices]

Hi! This is the friendly automated conda-forge-webservice.

I just wanted to let you know that I started rerendering the recipe in #226.

[trivialfis]

@jakirkham I updated the R test file:

require(xgboost)
data(agaricus.train, package = "xgboost")
data(agaricus.test, package = "xgboost")
train <- agaricus.train
test <- agaricus.test
str(train)
dim(train$data)
dim(test$data)
class(train$data)[1]
class(train$label)

bst_sparse <- xgboost(
  x = train$data, y = as.factor(train$label),
  max_depth = 2,
  learning_rate = 1.0,
  nthread = 2,
  nrounds = 2,
  objective = "binary:logistic"
)
bst_dense <- xgboost(
  x = as.matrix(train$data), y = as.factor(train$label),
  max_depth = 2,
  learning_rate = 1,
  nthread = 2,
  nrounds = 2,
  objective = "binary:logistic"
)

dtrain <- xgb.DMatrix(data = train$data, label = train$label)
bst_dmatrix <- xgb.train(
  params = xgb.params(
    max_depth = 2, learning_rate = 1, nthread = 2, objective = "binary:logistic"
  ),
  data = dtrain,
  nrounds = 2
)

pred <- predict(bst_sparse, test$data)
print(length(pred))
print(head(pred))

prediction <- as.numeric(pred > 0.5)
print(head(prediction))
err0 <- mean(as.numeric(pred > 0.5) != test$label)
print(paste("test-error=", err0))

dtrain <- xgb.DMatrix(data = train$data, label = train$label)
dtest <- xgb.DMatrix(data = test$data, label = test$label)
evals <- list(train = dtrain, test = dtest)
bst <- xgb.train(
  params = xgb.params(
    max_depth = 2, learning_rate = 1, nthread = 2, objective = "binary:logistic"
  ),
  nrounds = 2,
  evals = evals,
  data = dtrain
)


pred <- predict(bst, test$data)
print(length(pred))
print(head(pred))

prediction <- as.numeric(pred > 0.5)
print(head(prediction))
err1 <- mean(as.numeric(pred > 0.5) != test$label)
print(paste("test-error=", err1))

if (abs((err0 - err1)) > 1e-5) {
  stop()
}

Feel free to merge it into your PR.

[mzjp2]

Hey @jakirkham -- thanks a lot for the work you've done on this so far. Just wondering what the current state of the build is and what your next steps are. Anyway I/we can help?