Remove nccl_cvars.(h|cc) in favor of genrule-based generation

Ben Carver · facebook-github-bot · commit 132aa5f96209 · 2025-12-09T10:07:45.000-08:00
Summary:
This diff removes the `nccl_cvars.cc` and `nccl_cvars.h` files now that they'll instead be built/generated/provided by `genrule`.

This diff also updates the MCCL build script, `comms/mccl/build/build.sh`, to be compatible with the new `genrule`-based build for the `ncclx-cvars` library.

**Note**: this diff was originally published, landed, and reverted as D87668052.

Differential Revision: D88748045
diff --git a/build_ncclx.sh b/build_ncclx.sh
@@ -248,6 +248,41 @@ if [[ -z "${NCCL_BUILD_SKIP_DEPS}" ]]; then
   build_comms_tracing_service
 fi
 
+# Generate nccl_cvars files (these are no longer checked into the repo)
+# The files are generated by extractcvars.py which reads nccl_cvars.yaml and nccl_cvars.cc.in
+echo "Generating nccl_cvars files..."
+CVARS_DIR="$BASE_DIR/comms/utils/cvars"
+
+# Validate that the required source files exist
+if [ ! -f "$CVARS_DIR/extractcvars.py" ]; then
+  echo "ERROR: extractcvars.py not found at $CVARS_DIR/extractcvars.py"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.yaml" ]; then
+  echo "ERROR: nccl_cvars.yaml not found at $CVARS_DIR/nccl_cvars.yaml"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.cc.in" ]; then
+  echo "ERROR: nccl_cvars.cc.in not found at $CVARS_DIR/nccl_cvars.cc.in"
+  exit 1
+fi
+
+# Install ruamel-yaml if not already installed (required by extractcvars.py)
+if [[ -z "${NCCL_SKIP_CONDA_INSTALL}" ]]; then
+  conda install ruamel.yaml --yes
+fi
+
+# Run the extractcvars.py script directly to generate the files
+export NCCL_CVARS_OUTPUT_DIR="$CVARS_DIR"
+python3 "$CVARS_DIR/extractcvars.py"
+
+# Verify the files were generated
+if [ ! -f "$CVARS_DIR/nccl_cvars.h" ] || [ ! -f "$CVARS_DIR/nccl_cvars.cc" ]; then
+  echo "ERROR: Failed to generate nccl_cvars files"
+  exit 1
+fi
+echo "Successfully generated nccl_cvars files in $CVARS_DIR"
+
 # set up the third-party ldflags
 export PKG_CONFIG_PATH="${CONDA_LIB_DIR}"/pkgconfig
 THRIFT_SERVICE_LDFLAGS=(
diff --git a/build_rcclx.sh b/build_rcclx.sh
@@ -216,6 +216,41 @@ if [[ -z "${NCCL_BUILD_SKIP_DEPS}" ]]; then
   build_third_party
 fi
 
+# Generate nccl_cvars files (these are no longer checked into the repo)
+# The files are generated by extractcvars.py which reads nccl_cvars.yaml and nccl_cvars.cc.in
+echo "Generating nccl_cvars files..."
+CVARS_DIR="$BASE_DIR/comms/utils/cvars"
+
+# Validate that the required source files exist
+if [ ! -f "$CVARS_DIR/extractcvars.py" ]; then
+  echo "ERROR: extractcvars.py not found at $CVARS_DIR/extractcvars.py"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.yaml" ]; then
+  echo "ERROR: nccl_cvars.yaml not found at $CVARS_DIR/nccl_cvars.yaml"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.cc.in" ]; then
+  echo "ERROR: nccl_cvars.cc.in not found at $CVARS_DIR/nccl_cvars.cc.in"
+  exit 1
+fi
+
+# Install ruamel-yaml if not already installed (required by extractcvars.py)
+if [[ -z "${NCCL_SKIP_CONDA_INSTALL}" ]]; then
+  conda install ruamel.yaml --yes
+fi
+
+# Run the extractcvars.py script directly to generate the files
+export NCCL_CVARS_OUTPUT_DIR="$CVARS_DIR"
+python3 "$CVARS_DIR/extractcvars.py"
+
+# Verify the files were generated
+if [ ! -f "$CVARS_DIR/nccl_cvars.h" ] || [ ! -f "$CVARS_DIR/nccl_cvars.cc" ]; then
+  echo "ERROR: Failed to generate nccl_cvars files"
+  exit 1
+fi
+echo "Successfully generated nccl_cvars files in $CVARS_DIR"
+
 if [ "$CLEAN_BUILD" == 1 ]; then
     rm -rf "$BUILDDIR"
 fi
diff --git a/comms/ncclx/v2_27/maint/oss_build.sh b/comms/ncclx/v2_27/maint/oss_build.sh
@@ -288,6 +288,41 @@ fi
 
 mkdir -p $BUILDDIR
 
+# Generate nccl_cvars files (these are no longer checked into the repo)
+# The files are generated by extractcvars.py which reads nccl_cvars.yaml and nccl_cvars.cc.in
+echo "Generating nccl_cvars files..."
+CVARS_DIR="$FBCODE_DIR/comms/utils/cvars"
+
+# Validate that the required source files exist
+if [ ! -f "$CVARS_DIR/extractcvars.py" ]; then
+  echo "ERROR: extractcvars.py not found at $CVARS_DIR/extractcvars.py"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.yaml" ]; then
+  echo "ERROR: nccl_cvars.yaml not found at $CVARS_DIR/nccl_cvars.yaml"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.cc.in" ]; then
+  echo "ERROR: nccl_cvars.cc.in not found at $CVARS_DIR/nccl_cvars.cc.in"
+  exit 1
+fi
+
+# Install ruamel-yaml if not already installed (required by extractcvars.py)
+if [ -z "$SKIP_CONDA_INSTALL" ]; then
+  conda install -p "$CONDA_DIR" ruamel.yaml --yes
+fi
+
+# Run the extractcvars.py script directly to generate the files
+export NCCL_CVARS_OUTPUT_DIR="$CVARS_DIR"
+python3 "$CVARS_DIR/extractcvars.py"
+
+# Verify the files were generated
+if [ ! -f "$CVARS_DIR/nccl_cvars.h" ] || [ ! -f "$CVARS_DIR/nccl_cvars.cc" ]; then
+  echo "ERROR: Failed to generate nccl_cvars files"
+  exit 1
+fi
+echo "Successfully generated nccl_cvars files in $CVARS_DIR"
+
 # Use nccl relative to fbcode dir (configurable for Docker builds)
 export NCCL_HOME=${NCCL_HOME:-$FBCODE_DIR/comms/ncclx/v2_27}
 pushd "${NCCL_HOME}"
diff --git a/comms/utils/cvars/README.md b/comms/utils/cvars/README.md
@@ -1,53 +1,115 @@
 # Custom VARS
 
-NCCLX CVARS - Strongly typed configurable knobs for NCCLX. All
-configuration knobs are defined here, and can be used in source
-file by including `nccl_cvars.h` and use typed CVAR by its name.
+NCCLX CVARS - Strongly typed configurable knobs for NCCLX. All configuration
+knobs are defined here, and can be used in source file by including
+`nccl_cvars.h` and use typed CVAR by its name.
 
 ## User Guide
 
-Refer to `nccl_cvars.yaml` for CVAR documentation and their default
-values.
+Refer to `nccl_cvars.yaml` for CVAR documentation and their default values.
 
 CVAR can be provided to program in two ways
-1) Environment variable - e.g. `NCCL_DEBUG=warn nccl_allreduce_perf ...`
-2) Config Variable - define in `/etc/nccl.conf` and it'll be picked up
+
+1. Environment variable - e.g. `NCCL_DEBUG=warn nccl_allreduce_perf ...`
+2. Config Variable - define in `/etc/nccl.conf` and it'll be picked up
    automatically by program.
 
-Environment variable will take precedence over config variable. If not
-specified in either, the default value will be used.
+Environment variable will take precedence over config variable. If not specified
+in either, the default value will be used.
 
 ## Developer Guide
 
 All CVARs are defined in `nccl_cvars.yaml`. To add a new CVAR:
-1) Add the CVAR definition in `nccl_cvars.yaml`
-2) Build any target that depends on `//comms/utils/cvars:ncclx-cvars` - the files will be auto-generated via genrule
-3) Include `#include "comms/utils/cvars/nccl_cvars.h"` and use your CVAR in program
 
-**Note:** `nccl_cvars.h` and `nccl_cvars.cc` are now generated at build time using a genrule.
-They should **NOT** be manually edited or committed to the repository. The genrule automatically
-generates these files from `nccl_cvars.yaml` using `extractcvars.py` whenever you build a target
-that depends on the `ncclx-cvars` library.
+1. Add the CVAR definition in `nccl_cvars.yaml`
+2. Build any target that depends on `//comms/utils/cvars:ncclx-cvars` - the
+   files will be auto-generated via genrule
+3. Include `#include "comms/utils/cvars/nccl_cvars.h"` and use your CVAR in
+   program
+
+**Note:** `nccl_cvars.h` and `nccl_cvars.cc` are now generated at build time
+using a genrule. They should **NOT** be manually edited or committed to the
+repository. The genrule automatically generates these files from
+`nccl_cvars.yaml` using `extractcvars.py` whenever you build a target that
+depends on the `ncclx-cvars` library.
 
 To regenerate the files manually (for development/testing), you can run:
+
 ```bash
 cd ~/fbsource/fbcode && buck2 run comms/utils/cvars:extractcvars
 ```
 
-The CVAR is initialized as part of ncclInit and it is done by `initEnv` from `init.cc`. CVAR
-must not be used before initialization.
+The CVAR is initialized as part of ncclInit and it is done by `initEnv` from
+`init.cc`. CVAR must not be used before initialization.
+
+## Including CVARs in Build Scripts (OSS/Non-Buck Builds)
+
+For OSS builds or other build scripts that don't use Buck2, you need to generate
+the `nccl_cvars.h` and `nccl_cvars.cc` files before building. There are two
+approaches:
+
+### Option 1: Using Buck2 Genrule (when Buck2 is available)
+
+If Buck2 is available in your build environment, you can use the genrule to
+generate the files:
+
+```bash
+GENRULE_OUTPUT=$(buck2 build fbcode//comms/utils/cvars:generate_nccl_cvars --show-full-output 2>&1 | grep "generate_nccl_cvars" | awk '{print $2}')
+if [ -n "$GENRULE_OUTPUT" ]; then
+  cp "$GENRULE_OUTPUT/nccl_cvars.h" "$CVARS_DIR/nccl_cvars.h"
+  cp "$GENRULE_OUTPUT/nccl_cvars.cc" "$CVARS_DIR/nccl_cvars.cc"
+fi
+```
+
+### Option 2: Running extractcvars.py Directly (recommended for OSS builds)
+
+For builds outside of Buck2 (e.g., conda/Docker builds), run the
+`extractcvars.py` script directly:
+
+```bash
+# Set the output directory for the generated files
+CVARS_DIR="$FBCODE_DIR/comms/utils/cvars"
+
+# Validate that the required source files exist
+if [ ! -f "$CVARS_DIR/extractcvars.py" ]; then
+  echo "ERROR: extractcvars.py not found"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.yaml" ]; then
+  echo "ERROR: nccl_cvars.yaml not found"
+  exit 1
+fi
+if [ ! -f "$CVARS_DIR/nccl_cvars.cc.in" ]; then
+  echo "ERROR: nccl_cvars.cc.in not found"
+  exit 1
+fi
+
+# Install ruamel-yaml (required by extractcvars.py)
+conda install ruamel.yaml --yes  # or: pip install ruamel.yaml
+
+# Run the script to generate the files
+export NCCL_CVARS_OUTPUT_DIR="$CVARS_DIR"
+python3 "$CVARS_DIR/extractcvars.py"
+
+# Verify the files were generated
+if [ ! -f "$CVARS_DIR/nccl_cvars.h" ] || [ ! -f "$CVARS_DIR/nccl_cvars.cc" ]; then
+  echo "ERROR: Failed to generate nccl_cvars files"
+  exit 1
+fi
+```
 
 ## Changed NCCL CVAR Default values
 
-NCCL_RAS_ENABLE - default value changed from 1 to 0
-NCCL_CTRAN_IB_MAX_QPS - default value changed from 1 to 16
-NCCL_CTRAN_IB_QP_MAX_MSGS - default value changed from 4 to 128
-NCCL_CTRAN_IB_QP_SCALING_THRESHOLD - default value changed from 131072 to 524288
-NCCL_CTRAN_IB_QP_CONFIG_XDC - default value changed from "" to "1048576,16,spray,128"
-NCCL_CTRAN_IB_QP_CONFIG_XRACK - default value changed from "" to "1048576,16,spray,128"
-NCCL_CTRAN_IB_QP_CONFIG_XZONE - default value changed from "" to "1048576,16,spray,128"
-NCCL_CTRAN_IB_VC_MODE - default value changed from "spray" to "dqplb"
+NCCL_RAS_ENABLE - default value changed from 1 to 0 NCCL_CTRAN_IB_MAX_QPS -
+default value changed from 1 to 16 NCCL_CTRAN_IB_QP_MAX_MSGS - default value
+changed from 4 to 128 NCCL_CTRAN_IB_QP_SCALING_THRESHOLD - default value changed
+from 131072 to 524288 NCCL_CTRAN_IB_QP_CONFIG_XDC - default value changed from
+"" to "1048576,16,spray,128" NCCL_CTRAN_IB_QP_CONFIG_XRACK - default value
+changed from "" to "1048576,16,spray,128" NCCL_CTRAN_IB_QP_CONFIG_XZONE -
+default value changed from "" to "1048576,16,spray,128" NCCL_CTRAN_IB_VC_MODE -
+default value changed from "spray" to "dqplb"
 
 ## NCCL Baseline Adapter
 
-The NCCL Baseline Adapter API is designed to provide a similar interface to the baseline/third-party NCCL library's `ncclGetEnv` and `ncclLoadParam` functions.
+The NCCL Baseline Adapter API is designed to provide a similar interface to the
+baseline/third-party NCCL library's `ncclGetEnv` and `ncclLoadParam` functions.
diff --git a/comms/utils/cvars/nccl_cvars.cc b/comms/utils/cvars/nccl_cvars.cc
diff --git a/comms/utils/cvars/nccl_cvars.h b/comms/utils/cvars/nccl_cvars.h
