Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
107 changes: 100 additions & 7 deletions docker-bake.hcl
Original file line number Diff line number Diff line change
Expand Up @@ -52,6 +52,10 @@ variable "FFMPEG_VLM_WHEEL_CONTEXT" {
default = ""
}

variable "TRANSFORMER_ENGINE_WHEEL_CONTEXT" {
default = ""
}

variable "DISTROLESS_BASE" {
default = "nvcr.io/nvidia/distroless/python:3.11-v4.0.8"
}
Expand All @@ -78,6 +82,30 @@ variable "BASE_TAG_AUTOMODEL" {
default = "2c1a9ef2535a6648a272d9b74dae97fb672b8234"
}

# Pin for nmp-rl-base (prebuilt-base tag) + the NeMo-RL / NeMo-Gym source the base builds from.
# Defaults point at the soluwalana/{RL,Gym} forks. If the RL fork's uv.lock and the Gym ref drift,
# build with UV_SYNC_MODE= to relock.
variable "BASE_TAG_RL" {
default = "local"
}
variable "NEMO_RL_REPO" {
default = "https://github.com/soluwalana/RL.git"
}
# Pin to an immutable commit SHA, not the branch name. The base clones RL at this ref
# (docker/rl/Dockerfile.nmp-rl-base, nemo-rl stage); a branch ref would let the fork move under us,
# silently invalidating the base's heavy uv-sync layer on every training rebuild (and breaking
# `uv sync --frozen` if that commit's lock drifted). Bump deliberately, in lockstep with the fork's
# uv.lock, when advancing RL.
#
# RL pins Gym as a git submodule (-> soluwalana/Gym over https), so Gym rides in with the RL git ADD
# - no separate Gym pin needed.
variable "NEMO_RL_REF" {
default = "14c0e020aa0c2c8ab55f4c57e7cea645ae6eb373" # soluwalana/RL nmp/customizer
}
variable "RL_BASE_CONTEXT" {
default = ""
}

# The tag for base images if needed
variable "WHEELS_TAG" {
default = "2c1a9ef2535a6648a272d9b74dae97fb672b8234"
Expand Down Expand Up @@ -143,6 +171,11 @@ function "get_ffmpeg_vlm_wheel_image" {
result = "${WHEELS_REGISTRY}/ffmpeg-vlm-wheel:${WHEELS_TAG}"
}

function "get_transformer_engine_wheel_image" {
params = []
result = "${WHEELS_REGISTRY}/transformer-engine-wheel:${WHEELS_TAG}"
}

function "get_arch_tag" {
params = []
result = BUILD_ARCH == "linux/arm64" ? "linux-arm64" : "linux-amd64"
Expand All @@ -160,6 +193,11 @@ function "automodel_base_context" {
result = notequal(AUTOMODEL_BASE_CONTEXT, "") ? AUTOMODEL_BASE_CONTEXT : notequal(USE_PREBUILT_BASES, "") ? "docker-image://${BASE_REGISTRY}/nmp-automodel-base:${BASE_TAG_AUTOMODEL}" : "target:nmp-automodel-base-builder"
}

function "rl_base_context" {
params = []
result = notequal(RL_BASE_CONTEXT, "") ? RL_BASE_CONTEXT : notequal(USE_PREBUILT_BASES, "") ? "docker-image://${BASE_REGISTRY}/nmp-rl-base:${BASE_TAG_RL}" : "target:nmp-rl-base-builder"
}

function "causal_conv1d_wheel_context" {
params = []
result = notequal(CAUSAL_CONV1D_WHEEL_CONTEXT, "") ? CAUSAL_CONV1D_WHEEL_CONTEXT : notequal(USE_LOCAL_WHEELS, "") ? "target:causal-conv1d-wheel" : "docker-image://${get_causal_conv1d_wheel_image()}"
Expand All @@ -175,6 +213,11 @@ function "ffmpeg_vlm_wheel_context" {
result = notequal(FFMPEG_VLM_WHEEL_CONTEXT, "") ? FFMPEG_VLM_WHEEL_CONTEXT : notequal(USE_LOCAL_WHEELS, "") ? "target:ffmpeg-vlm-wheel" : "docker-image://${get_ffmpeg_vlm_wheel_image()}"
}

function "transformer_engine_wheel_context" {
params = []
result = notequal(TRANSFORMER_ENGINE_WHEEL_CONTEXT, "") ? TRANSFORMER_ENGINE_WHEEL_CONTEXT : notequal(USE_LOCAL_WHEELS, "") ? "target:transformer-engine-wheel" : "docker-image://${get_transformer_engine_wheel_image()}"
}

function "wheel_tags" {
params = [name]
result = [
Expand Down Expand Up @@ -316,6 +359,7 @@ group "nmp-rl" {
targets = [
"nmp-rl-base-builder",
"nmp-rl-training",
"nmp-rl-training-smoke-test",
"nmp-customizer-tasks",
]
}
Expand Down Expand Up @@ -380,27 +424,36 @@ target "rl-platform-workspace" {
platforms = get_platforms()
}

# Heavy base: NGC torch/CUDA + NeMo-RL v0.4.0 + Ray.
# Heavy base: cuda-dl-base + NeMo-RL (with its Gym/Automodel/Megatron-Bridge submodules) built FROM
# SOURCE (Python 3.13, CUDA 13). RL is pinned via NEMO_RL_REF; Gym rides in as RL's own submodule
# (no separate pin). The ffmpeg-vlm wheels are reused for the cp313 av/opencv/decord2 CVE swap.
target "nmp-rl-base-builder" {
target = "nmp-rl-base"
context = "."
dockerfile = "docker/Dockerfile.nmp-rl-base"
dockerfile = "docker/rl/Dockerfile.nmp-rl-base"
contexts = {
ffmpeg-vlm-wheel-image = ffmpeg_vlm_wheel_context()
}
args = {
NEMO_RL_REPO = NEMO_RL_REPO
NEMO_RL_REF = NEMO_RL_REF
}
cache-to = maybe_registry_cache_to("nmp-rl-base")
cache-from = maybe_registry_cache_from("nmp-rl-base")
tags = base_tags("nmp-rl-base")
output = image_output()
platforms = get_platforms()
}

# GPU DPO training image: base + platform glue. Bootstraps Ray at runtime.
# GPU DPO + GRPO training image (also the Gym environment runtime): base + platform glue.
# Bootstraps Ray at runtime.
target "nmp-rl-training" {
target = "runtime"
context = "."
dockerfile = "docker/Dockerfile.nmp-rl-training"
dockerfile = "docker/rl/Dockerfile.nmp-rl-training"
contexts = {
platform-workspace = "target:rl-platform-workspace"
nmp-rl-base = "target:nmp-rl-base-builder"
ffmpeg-vlm-wheel-image = ffmpeg_vlm_wheel_context()
platform-workspace = "target:rl-platform-workspace"
nmp-rl-base = rl_base_context()
}
cache-to = maybe_registry_cache_to("nmp-rl-training")
cache-from = maybe_registry_cache_from("nmp-rl-training")
Expand All @@ -409,6 +462,23 @@ target "nmp-rl-training" {
platforms = get_platforms()
}

# CPU-only import smoke tests for the training image (smoke-test stage runs pytest during build).
target "nmp-rl-training-smoke-test" {
target = "smoke-test"
context = "."
dockerfile = "docker/rl/Dockerfile.nmp-rl-training"
contexts = {
platform-workspace = "target:rl-platform-workspace"
nmp-rl-base = rl_base_context()
}
args = {
SMOKE_MARKER = "smoke_nmp_rl_training"
}
cache-from = maybe_registry_cache_from("nmp-rl-training")
output = ["type=cacheonly"]
platforms = get_platforms()
}

# Base images for consolidated containers
target "nmp-python-base" {
target = "nmp-python-base-builder"
Expand Down Expand Up @@ -639,6 +709,29 @@ target "ffmpeg-vlm-wheel" {
platforms = get_platforms()
}

# transformer-engine wheel (cp313 / cu130), pinned to NeMo-RL's TE ref (release_v2.15, from RL's
# override-dependencies) and arch (90;100, matching RL).
#
# NOT consumed by any image today, and deliberately not in a build group: Transformer-Engine only
# ships in the `automodel` and `mcore` extras, and the customizer's DPO + GRPO path uses neither
# (DPO -> fsdp, GRPO generation -> vllm). Kept because it builds cleanly and is the drop-in for
# nmp-rl-base if a Megatron/automodel backend is ever adopted. Build on demand:
# docker buildx bake transformer-engine-wheel
target "transformer-engine-wheel" {
target = "transformer-engine-wheel"
context = "."
dockerfile = "docker/base/Dockerfile.python-wheels"
cache-to = maybe_registry_cache_to("transformer-engine-wheel")
cache-from = maybe_registry_cache_from("transformer-engine-wheel")
tags = wheel_tags("transformer-engine-wheel")
output = image_output()
args = {
TE_REF = "release_v2.15"
NVTE_CUDA_ARCHS = "90;100"
}
platforms = get_platforms()
}


target "safe-synthesizer-tasks-docker" {
target = "runtime"
Expand Down
56 changes: 0 additions & 56 deletions docker/Dockerfile.nmp-rl-base

This file was deleted.

142 changes: 0 additions & 142 deletions docker/Dockerfile.nmp-rl-training

This file was deleted.

Loading
Loading