specflux 0.0.1

A JAX-only neural-network library with a PyTorch-shaped eager surface and Flax-NNX-style graph/state underneath.

specflux

True MPMD pipeline parallelism for JAX — with an eager module API.

Quick Start | Installation | MPMD Runtime | Examples | Docs

---

SpecFlux is a JAX-native neural network library built around true MPMD pipeline parallelism. Each physical rank compiles and runs its own XLA program — no shared shard_map HLO, no SPMD-same-shape constraint. Heterogeneous stages (embed → blocks → head), multiple schedules (GPipe, 1F1B, ZeroBubble, Interleaved, DualPipe), and a unified spx.run() entry point that dispatches to SPMD or MPMD from the same training script.

The module API is eager and debuggable — subclass Module, override forward, call model(x) — but every Module is a JAX pytree, so jax.jit, jax.grad, and jax.tree.map work directly.

---

Why specflux?

Capability	specflux	JAX + manual	other JAX frameworks
True MPMD	built-in (sxcall, sxjit)	hand-rolled	SPMD-only
Heterogeneous stages	native (different class/shape per rank)	fragile	not supported
Pipeline schedules	9 schedules (GPipe→DualPipeV)	hand-rolled	limited
Unified runtime	spx.run(mesh) → SPMD or MPMD	separate code paths	separate APIs
Eager modules	model(x) + pytree-native	functional only	functional or ref-tracking
Dispatch overhead	~150 µs	~N/A	~300–2000 µs

Dispatch overhead measured on a tiny 2-layer CPU transformer. See benchmarks.

---

Quick Start

pip install specflux

Single-device eager training

import jax.numpy as jnp
import specflux as spx
from specflux import nn

class MLP(spx.Module):
    def __init__(self, d, h, o, *, rngs):
        super().__init__()
        self.fc1 = nn.Linear(d, h, rngs=rngs)
        self.fc2 = nn.Linear(h, o, rngs=rngs)

    def forward(self, x):
        return self.fc2(nn.gelu(self.fc1(x)))

model = MLP(16, 64, 4, rngs=spx.Rngs(0))

@spx.jit
@spx.value_and_grad
def loss_fn(m, x, y):
    return ((m(x) - y) ** 2).mean()

loss, grads = loss_fn(model, jnp.ones((8, 16)), jnp.zeros((8, 4)))

Marker-based MPMD — split a function into per-rank programs

from specflux.runtime.mpmd import sxjit, sxstage_iter
from specflux.pipeline import Std1F1B

# Define a multi-stage forward with explicit stage boundaries
@sxjit(schedule=Std1F1B(microbatches=8))
def pipeline_fwd(model, x):
    x = model.embed(x)                # stage 0
    x = sxstage_iter(x)               # boundary — split here
    x = model.blocks[0](x)            # stage 1
    x = sxstage_iter(x)               # boundary — split here
    x = model.blocks[1](x)            # stage 2
    x = sxstage_iter(x)               # boundary — split here
    return model.head(x)              # stage 3

# sxjit traces the function, splits the jaxpr at sxstage_iter markers,
# and compiles one XLA executable per stage/rank. Each rank only sees
# its own sub-graph — true MPMD, not SPMD.
output = pipeline_fwd(model, x)

MPMD pipeline training — one call, multiple devices

from specflux.sharding import logical_axis_rules

# Create a 4-stage pipeline mesh
mesh = spx.create_mesh(axis_dims=(2, 1, -1, 1, 1, 1), mpmd_axis="pp")

with logical_axis_rules(FSDP_TP_RULES), mesh:
    # MPMD: each rank gets its own compiled stage
    loss, grads = spx.run(
        model, inputs=x, targets=y,
        mesh=mesh, mode="train", loss_fn=ce_loss, microbatches=8,
    )

# Drop mpmd_axis → same code runs under pure SPMD pjit
# Same model, same script, different mesh — no code changes.

Deferred initialization — infer shapes at runtime

model = nn.Sequential(
    nn.Linear(None, 256, rngs=rngs),   # in_features inferred from first call
    nn.ReLU(),
    nn.Linear(256, 10, rngs=rngs),
)
y = model(jnp.zeros((8, 128)))         # weight shapes resolved here

---

MPMD Runtime

SpecFlux implements true MPMD: each physical rank compiles and executes its own distinct JAX program. This is not SPMD-with-barriers — stages can have different classes, different parameter shapes, and different I/O shapes.

spx.run — unified entry point

spx.run routes to SPMD (pjit) or MPMD (sxcall) based on the mesh:

spx.run(
    model,
    inputs=x,                # microbatched along leading axis
    targets=y,
    mesh=mesh,               # SpxMesh — mpmd_axis decides the path
    mode="train",            # "forward" | "train"
    loss_fn=ce_loss,
    microbatches=8,
)

Mesh type	What happens
mpmd_axis=None	Pure SPMD — pjit with FSDP/TP via logical axis rules
mpmd_axis="pp"	True MPMD — auto-split into stages, per-rank compilation

Lower-level primitives

For full control, drop below spx.run:

Primitive	Purpose
sxcall	Execute a PipelineSequential under a schedule — Python dispatch loop over pre-built per-rank callables
sxjit	Decorator: trace a function, split at sxstage_iter markers, compile one XLA executable per stage/rank
sxgrad / sxvalue_and_grad	Schedule-faithful gradients of an sxjit function
treduce	Schedule-driven microbatch reduction primitive — binds a body + schedule into the traced jaxpr
sxstage_iter	Marker primitive for stage boundaries inside sxjit

Supported schedules

Schedule	Type	Key trait
GPipe	Flat	All forwards, then all backwards. Simple, high memory.
Std1F1B	Flat	Standard 1-forward-1-backward. Peak memory O(n_stages).
ZeroBubbleH1	Flat	Splits BWD into input-grad + weight-grad; weight-grad fills bubble slots.
InterleavedH1	Virtual	Each rank owns v non-contiguous stages. Bubble shrinks by v.
DualPipeV	Virtual	V-shaped bidirectional pipeline (DeepSeek-style).
KimiK2	Virtual	Interleaved with extra warmup (Moonshot K2 design).

Heterogeneous stages

No same-shape constraint. Stages can be completely different:

model = PipelineSequential(
    EmbedStage(vocab, d, rngs=rngs),       # (B, S) int → (B, S, d)
    BlockStage(d, rngs=rngs),               # (B, S, d) → (B, S, d)
    BlockStage(d, rngs=rngs),               # (B, S, d) → (B, S, d)
    HeadStage(d, vocab, rngs=rngs),        # (B, S, d) → (B, S, vocab)
)

Auto-splitting is available for homogeneous stacks: spx.run detects model.blocks: ModuleList and slices it evenly across pipeline stages.

---

Installation

pip install specflux

# Optional extras
pip install "specflux[contrib]"   # optax integration
pip install "specflux[cuda]"      # CUDA jaxlib
pip install "specflux[tpu]"       # TPU jaxlib

From source:

uv sync --extra dev --extra test --extra contrib

Requires Python 3.11+ and JAX >= 0.9.2.

---

Features

MPMD Pipeline Parallelism

True multi-program multi-data execution with per-rank compilation, schedule-faithful dispatch, and heterogeneous stage support.

Module-aware JAX transforms

Transform	What it does
spx.jit	mutable= selector declares writable collections
spx.grad / spx.value_and_grad	wrt= selector picks the differentiated subset
spx.vmap	module states passed with in_axes=None automatically
spx.scan / spx.remat_scan	module-aware loops, optionally checkpointed
spx.remat	gradient checkpointing
spx.cond / spx.switch / spx.while_loop / spx.fori_loop	control flow over module state

Selector DSL

One composable predicate for every "subset of the model" API:

spx.grad(loss, wrt="params")                     # by collection name
spx.grad(loss, wrt=nn.LoraParameter)             # by Variable class
spx.grad(loss, wrt=spx.path_contains("attn"))    # by path glob

sel = spx.select().at_instances_of(nn.Linear).of_type(spx.Parameter) - spx.path_contains("head")
trainable, frozen = sel.partition_state(model, state)

Sharding & SPMD

Annotate variables with logical axis names and let the mesh decide:

w = spx.Parameter(
    jnp.zeros((256, 256)),
    sharding=spx.Sharding(("data", "model")),
    axis_names=("in", "out"),
)

LoRA Fine-Tuning

base = nn.Linear(768, 768, rngs=spx.Rngs(0))
model = nn.wrap_lora(base, rank=8, alpha=16, rngs=spx.Rngs(1))

@spx.jit
@spx.grad(wrt="lora")
def step(m, x, y):
    return ((m(x) - y) ** 2).mean()

FP8 Training

Delayed-scaling per-tensor FP8 with rolling amax history:

@spx.jit(mutable="fp8_meta")
def step(m, x, y):
    def loss(m, x, y):
        return ((m(x) - y) ** 2).mean()
    return spx.grad(loss)(m, x, y)

Explicit graph / state seam

gdef, state = spx.export(model)          # immutable GraphDef + State dict
model2 = spx.bind(gdef, state)           # reconstruct, skips __init__
spx.update(model, state)                 # in-place state patch
clone = spx.clone(model)                 # deep-copy via export+bind

Inspection

spx.inspect.summary(model, jnp.zeros((1, 128)))
spx.inspect.count_params(model)
spx.inspect.count_bytes(model)
spx.inspect.tabulate(model, example_input)

---

Examples

See examples/ for runnable scripts:

Folder	Topic
01_basics/	Modules, training loops, export/bind, optimizers
02_implementation_guide/	Llama 3, Qwen 2, GPT-2, ViT, custom transformer
03_transformations/	jit, grad, vmap, remat, scan, fori_loop
04_surgery/	Selectors, LoRA injection, FP8, freezing, swapping
05_shardings/	FSDP, TP, hybrid sharding, logical axis rules
06_spmd_scheduled/	Pipeline runtime with all schedules
07_mpmd/	Real MPMD pipeline via spx.run — train, forward, decode, 3-D mesh

python -m examples.01_basics.02_training_loop
python -m examples.02_implementation_guide.01_llama3
python -m examples.07_mpmd.01_train_homogeneous

---

Design

1. True MPMD first — sxcall compiles one XLA program per physical rank. Stages can differ in class, shape, and parameters. No SPMD-same-shape constraint.
2. Unified runtime — spx.run dispatches on the mesh. Same model, same script; change the mesh and you change the parallelism strategy.
3. Schedule-faithful execution — the dispatch loop walks the schedule grid exactly as planned. No hidden reordering, no implicit fusing.
4. Modules are JAX pytrees — flatten/unflatten through export/bind. jax.jit, jax.tree.map, jax.value_and_grad work directly.
5. State lives in Variable cells — Parameter, Buffer, custom subclasses. Each tags its collection (params, buffers, lora, fp8_meta, ...).
6. One filter DSL everywhere — Selector serves grad(wrt=...), jit(mutable=...), Optimizer(wrt=...), freeze(...), iter_variables(select=...).

---

Benchmarks

python -m benchmarks.bench --cases all --device cpu
python -m benchmarks.bench --cases large --device tpu     # 1.21B transformer

Results land in benchmarks/results/latest.{json,md} with per-case specflux_ms / nnx_ms ratios.

On a tiny CPU dispatch-bound benchmark (2-layer / d=64 / batch-2 transformer), specflux runs at 1.83× the speed of flax.nnx; on d=48 it hits 2.0×. On compute-bound workloads (TPU 8B) the Python gap shrinks but stays positive.

---

Testing

pytest tests/ -q
pytest tests/test_conformance.py

---

Status

v0.0.1 — alpha. API may still move; pin the version if you depend on behavioural stability.

---

License

Apache-2.0.