spectrax-lib 0.0.7

SpectraX: a JAX-only neural-network library with a True MPMD pipeline parallelism for JAX — with an eager module API.

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

Quick Start · Installation · MPMD Runtime · Sharding · Benchmark · Examples · Docs

---

SpectraX 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), nine pipeline schedules (GPipe, 1F1B, ZeroBubble, Interleaved, DualPipeV, …), and a unified spx.run()/spx.jit() 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.

Used in production. EasyDeL is built on SpectraX: 77 model families (Llama, DeepSeek-V3, Qwen3, Gemma, GPT-OSS, Mamba, Whisper, …) with shared sharding, KV-cache, and pipeline plumbing. SpectraX is the JAX-only NN core; EasyDeL is the model zoo + training/serving stack.

---

Why SpectraX?

Capability	SpectraX	JAX + manual	other JAX frameworks
True MPMD	built-in (sxcall, sxjit, spx.run)	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
Symbolic sharding	BATCH, EMBED, TP, FSDP, …	string-typed P	manual PartitionSpec
Dispatch overhead	~150 µs	~N/A	~300–2000 µs

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

---

Quick Start

pip install spectrax-lib

1 · Single-device eager training

import jax.numpy as jnp
import spectrax as spx
from spectrax 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)))

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

from spectrax.runtime.mpmd import sxjit, sxstage_iter
from spectrax.runtime.schedules import Std1F1B
from jax.sharding import PartitionSpec as P

mesh = spx.create_mesh(axis_dims=(4, 1, -1, 1, 1, 1), mpmd_axis="pp")

@sxjit(mesh=mesh, schedule=Std1F1B(microbatches=8))
def pipeline_fwd(model, x):
    x = model.embed(x)                                 # stage 0
    x = sxstage_iter(x, sharding=P("fsdp", None, "tp"))   # boundary — declares the
                                                       # activation sharding the
                                                       # next stage receives
    x = model.blocks[0](x)                             # stage 1
    x = sxstage_iter(x, sharding=P("fsdp", None, "tp"))   # same contract again
    x = model.blocks[1](x)                             # stage 2
    x = sxstage_iter(x)                                # plain identity boundary
    return model.head(x)                               # stage 3

sxjit traces, 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.

Each marker is functionally the identity but accepts two optional keywords:

Arg	Purpose
stage=	Integer hint for the stage index (validation / debugging only — the cluster splitter partitions purely by marker order).
sharding=	PartitionSpec declaring the activation layout that flows across this boundary. The compiler honors it during cross-rank transport.

This is the same pattern EasyDeL uses to keep activation sharding stable across stage boundaries — see easydel.infra.base_module._maybe_emit_stage_boundary.

3 · One-call MPMD training — spx.run

from spectrax.sharding import logical_axis_rules

mesh = spx.create_mesh(axis_dims=(2, 1, -1, 1, 1, 1), mpmd_axis="pp")

with logical_axis_rules(FSDP_TP_RULES), mesh:
    loss, grads = spx.run(
        model,
        inputs=ids,
        targets=labels,
        mesh=mesh,
        mode="train",
        loss_fn=cross_entropy,
        microbatches=8,
    )

Drop mpmd_axis="pp" and the same code path runs under pure SPMD — same model, same script, no rewrites.

4 · 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

SpectraX 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.

The public MPMD surface is intentionally narrow and honest: sxjit, sxgrad, sxvalue_and_grad, sxcall, spx.run with an MPMD mesh, and MpmdPipelineExecutor all route through the true MPMD runtime. SPMD-only helpers reject MPMD-tagged meshes instead of silently taking a shard-map or host-jit fallback.

spx.run — the 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,
    schedule=Std1F1B(...),   # optional; defaults to GPipe
)

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

For mode="train", inputs/targets accept arrays, tuples, or dicts; SpectraX threads them into model.forward(...) and loss_fn(...).

Lower-level primitives

Primitive	Purpose
sxcall	Execute a PipelineSequential through the same scheduled MPMD dispatcher used by sxjit.
sxjit	Decorator: trace -> 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 — binds a body + schedule into the traced jaxpr.
sxstage_iter	Marker primitive for stage boundaries inside sxjit.
sxstage_region	Region marker for multimodal/branched pipelines that need separate logical stage sequences.
spx.assign_stage	Context manager that stamps subsequently-created variables with a (current, total) stage tag.

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).
Eager1F1B	Flat	1F1B variant that aggressively starts the backward pipe.
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.
InterleavedGPipe	Virtual	GPipe analog of InterleavedH1.
Interleaved1F1BPlusOne	Virtual	Interleaved 1F1B with one extra warmup microbatch.
DualPipeV	Virtual	V-shaped bidirectional pipeline (DeepSeek-style).
KimiK2	Virtual	Interleaved with extra warmup (Moonshot K2 design).

Choosing & using a schedule

All schedules share the same constructor shape: microbatches= is required; virtual-stage schedules also take virtual_stages=. Every schedule object exposes bubble_ratio(n_stages), peak_activations(n_stages), and total_steps(n_stages) so you can compare them analytically before launching:

from spectrax.runtime.schedules import (
    GPipe, Std1F1B, ZeroBubbleH1, InterleavedH1, DualPipeV,
)

n = 4   # pipeline depth
m = 16  # microbatches

for sc in [
    GPipe(m),
    Std1F1B(m),
    ZeroBubbleH1(m),
    InterleavedH1(m, virtual_stages=2),
    DualPipeV(m),
]:
    print(f"{type(sc).__name__:14s}  bubble={sc.bubble_ratio(n):.3f}  "
          f"steps={sc.total_steps(n)}  peak_acts={sc.peak_activations(n)}")

Pass the chosen schedule to spx.run, spx.jit, sxcall, or sxjit:

# spx.run — picks SPMD or MPMD by mesh
loss, grads = spx.run(
    model, inputs=x, targets=y, mesh=mesh,
    mode="train", loss_fn=ce_loss,
    microbatches=m, schedule=Std1F1B(m),
)

# sxcall — explicit MPMD on a PipelineSequential
loss, grads = sxcall(
    model, (x, y),
    mesh=mpmd_mesh, schedule=ZeroBubbleH1(m), loss_fn=ce_loss,
)

# sxjit — marker-based MPMD compile
@sxjit(mesh=mesh, schedule=InterleavedH1(m, virtual_stages=2))
def step(model, x):
    ...

Legacy knobs that forced host-side schedule walking (fuse_1f1b, fuse_zb, chunks, non-device_put transports, activation donation) are rejected on the true scheduled MPMD path. Use schedules that emit the desired fused cells directly.

Rough rule of thumb:

Goal	Pick
Simplest schedule	GPipe
Steady-state memory O(n_stages)	Std1F1B
Fill the 1F1B bubble	ZeroBubbleH1
Shrink bubble further at extra transport	InterleavedH1(virtual_stages=v)
DeepSeek-style V-shape	DualPipeV
Long-context / Moonshot-K2-style	KimiK2

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)
)

spx.run also auto-splits homogeneous stacks: it detects a ModuleList named blocks and slices it evenly across pipeline stages — no pipeline-aware bookkeeping in user code.

---

Sharding: the symbolic-axis system

SpectraX separates what a tensor axis means (BATCH, EMBED, TP, FSDP) from which mesh axis it lands on. Layer authors write symbolic tokens; the runtime picks the actual PartitionSpec based on the active mesh and runtime mode (training vs. autoregressive decode).

from spectrax import PartitionAxis, PartitionManager, common_types as ct

paxis = PartitionAxis()  # defaults: dp/fsdp/tp/sp/ep
with PartitionManager(paxis):
    h = ct.apply_logical_sharding(h, dynamic_axes=ct.HiddenStateSharding)
    # h is now constrained to (BATCH, QUERY_LENGTH, EMBED)
    # → resolved at runtime to whatever mesh the user is running on

Pre-baked symbolic shapes ship for the common cases: HiddenStateSharding, AttnQSharding, AttnKVSharding, RowWise, ColumnWise, Replicated, plus an Expert* family for MoE.

This is the same axis system EasyDeL uses to ship 77 model families without per-model sharding code — see easydel.axis for an example of registering a domain-specific axis token (ATTN_DP) that shadows DP only for KV-cache placement.

---

Module-aware JAX transforms

Transform	What it does
spx.jit	mutable= + mesh= + schedule= — one entry, SPMD/MPMD
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 (function- and class-aware)
spx.cond / spx.switch / spx.while_loop / spx.fori_loop	module-aware control flow
spx.eval_shape	shape-inference without running ops
spx.jvp / spx.vjp	forward / reverse-mode lifts that respect mutable=

spx.jit in depth

spx.jit accepts every keyword jax.jit does, plus three SpectraX-only ones: mutable=, mesh=, and schedule=. The mesh decides where the call goes:

@spx.jit                                 # plain jax.jit — no mesh
def fwd(model, x): return model(x)

@spx.jit(mutable="batch_stats")          # writes to "batch_stats" survive the call
def train_bn(model, x):
    return ((model(x)) ** 2).mean()

@spx.jit(mesh=spmd_mesh)                 # SPMD pjit — uses spmd_mesh
def fwd(model, x): return model(x)

@spx.jit(mesh=mpmd_mesh, schedule=Std1F1B(microbatches=8))
def step(model, x):                      # MPMD — internally calls sxjit
    x = model.embed(x)
    x = sxstage_iter(x)
    x = model.blocks[0](x)
    ...

Keyword	Meaning
mutable=	Selector / collection-name sugar: which collections may be written back after the call. Anything else throws IllegalMutationError. Default () (read-only).
mesh=	SpxMesh. If mesh.is_mpmd, the call is forwarded to sxjit (true MPMD). Otherwise plain jax.jit.
schedule=	Schedule instance, only meaningful when mesh is MPMD. Forwarded to sxjit.
static_argnums / static_argnames	Same as jax.jit. Static Module args are recommended (graph baked into the trace).
donate_argnums	Same as jax.jit. Useful for handing off optimizer state in-place under MPMD.
in_shardings / out_shardings	Same as jax.jit. Forwarded verbatim under SPMD; under MPMD they are resolved against per-stage sub-meshes.
keep_unused, device, backend, inline, compiler_options	Forwarded to jax.jit. Rejected with a clear error when mesh is MPMD (no analog in sxjit).

When mutable= is set, the compiled function takes the form (states, stripped_args, stripped_kwargs) internally — static_argnums and donate_argnums index into that 3-tuple, so prefer the *_argnames variants whenever possible.

The cache is keyed on the input modules' GraphDef snapshot. Mutating parameter values does not invalidate the cache; structural changes (adding/removing/replacing a Module or Variable attribute) do.

Selector DSL

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

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

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

The same DSL backs grad(wrt=...), jit(mutable=...), Optimizer(wrt=...), freeze(...), and iter_variables(select=...).

---

Features

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, 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_parameters(model)
spx.inspect.count_bytes(model)
spx.inspect.tabulate(model, example_input)

Built-in layers

Linear, Conv1d/2d/3d + ConvTranspose*, MultiheadAttention, LayerNorm, RMSNorm, BatchNorm, InstanceNorm, GroupNorm, Dropout, Embed, MLPBlock, MoE primitives, RNN/GRU/LSTM cells, FP8 path, LoRA path, plus containers (Sequential, ModuleList, ModuleDict, StackedModuleList for repeated transformer blocks under lax.scan).

---

Installation

pip install spectrax-lib

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

From source:

git clone https://github.com/erfanzar/spectrax
cd spectrax
uv sync --extra dev --extra test --extra contrib

Requires Python 3.11+ and JAX ≥ 0.9.2.

---

Examples

Seven topic folders under examples/ progress from single-Module forward passes to multi-device MPMD pipeline training:

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

Most examples run on CPU with small configs; sharding and pipeline examples benefit from multi-device TPU / GPU but fall back cleanly to one device.

---

Used by

• EasyDeL — production training/serving framework for LLMs, multimodal models, and vision models. SpectraX is the NN core; EasyDeL adds the model zoo (Llama, Qwen3, DeepSeek-V3, Gemma, GPT-OSS, Mamba, Whisper, …), trainers, KV cache, and inference engine. EasyDeL leans on spx.Module, spx.Rngs, spx.Parameter, spx.assign_stage, spx.sxstage_iter, the PartitionAxis registry, and the common_types symbolic-axis tokens.

If your project uses SpectraX, open a PR to add it here.

---

Design

1. True MPMD first. sxcall and sxjit use the scheduled MPMD dispatcher, compiling distinct per-rank programs. 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 runtime follows the schedule grid exactly as planned. No hidden SPMD fallback, no implicit legacy schedule walker.
4. Modules are JAX pytrees. Flatten/unflatten via export/bind. jax.jit, jax.tree.map, jax.value_and_grad work directly.
5. State lives in Variable cells. Parameter, Buffer, and user subclasses; each tags its collection (parameters, buffers, lora, fp8_meta, …).
6. One filter DSL everywhere. Selector serves grad(wrt=...), jit(mutable=...), Optimizer(wrt=...), freeze(...), iter_variables(select=...).
7. Symbolic sharding tokens. Layer code says BATCH, EMBED, TP; PartitionManager resolves the active mesh at runtime. Decode-mode overrides flow through the same tokens — no per-mode forking in layer code.

---

Benchmark

python -m benchmarks.bench --cases all --device cpu
python -m benchmarks.bench --cases large --device tpu     # 1.21B transformer
python -m benchmarks.llama_transforms_3way --preset small --device tpu --plots

The CPU dispatch benchmark writes benchmarks/results/latest.{json,md} with per-case spectrax_ms / nnx_ms ratios. On a tiny CPU dispatch-bound benchmark (2-layer / d=64 / batch-2 transformer), SpectraX runs at 1.83x the speed of flax.nnx; on d=48 it hits 2.0x. On compute-bound workloads (TPU 8B) the Python gap shrinks but stays positive.

The plots below are from the TPU small three-way Llama transform benchmark, comparing raw JAX, Flax NNX, and SpectraX across the transform surface used by real training code: jit, grad, value_and_grad, jvp, vjp, vmap, control flow, scan, and remat. Ratios are runtime against raw JAX, so 1.0 is parity, lower is faster, higher is slower.

Compile latency is tracked separately because first-call XLA costs can dominate short benchmark cases and tell a different story from steady-state runtime.

---

Testing

pytest tests/ -q
pytest tests/test_conformance.py

---

Status

v0.0.3 — alpha. API may still move; pin the version if you depend on behavioural stability. See CHANGELOG.md for the release log.

---

License

This project is licensed under AGPL-3.0-or-later.

If you use EasyDeL or SpecTrax in research, infrastructure, benchmarks, or derivative systems, please provide attribution and cite the project.

Citation

@software{easydel,
  author = {Erfan Zare Chavoshi},
  title = {EasyDeL},
  year = {2023},
  url = {<https://github.com/erfanzar/EasyDeL}>
}