xbill9/gemma-4-E4B-it-inferentia2

Gemma-4-E4B-it on AWS Inferentia2 — Tensor-Parallel (TP=2)

Compiled AWS Neuron artifacts that run `google/gemma-4-E4B-it` coherently across both NeuronCores of a single Inferentia2 device (inf2.8xlarge), at ~32–34 tokens/sec, with greedy decode token-for-token identical to the CPU reference.

This is a runnable inference port, not a fine-tune — the weights are Google's, unmodified. What's new is the tensor-parallel compilation recipe: a way to shard Gemma-4-E4B across two 16 GB cores and keep the KV cache device-resident, which the vendor stack (optimum-neuron / NxD) cannot currently do for this model.

Companion: the smaller `E2B` fits a single core and uses a different (simpler) design. E4B is ~2× larger and does not fit one core — it requires this TP=2 build.

TL;DR

⚡ Update — on-device prefill (recommended build)

The original build seeds the prompt with a host-CPU prefill (~1.4–1.6 s to first token). A newer build runs prefill on-device via a single weight-sharing ModelBuilder trace — prefill and decode as two buckets of one resident model, bf16 sharded weights, on-device aliased KV:

Fastest path — prebuilt Docker image (bundles the compiled model + server):

docker run -d --device /dev/neuron0 --ipc=host -p 8080:8080 \
  xbill9/gemma4-optb-e4b:tp2-devprefill-512
curl -s localhost:8080/generate -d '{"prompt":"What is AWS Inferentia?","max_tokens":60}'

Serves OpenAI-compatible routes (/v1/chat/completions, /v1/completions) plus /generate, /health, /metrics. To build it yourself: MB_WDTYPE=bf16 KV_MAX=512 KV_BUCKET=128 MB_SAVE=mb_e4b_512_bf16.pt python tp_mb.py (compiles + saves the model), then serve with optb_server_mb.py / package with Dockerfile.mb.

Run on a single `inf2.xlarge` (¼ the price) — slim image. The inf2.xlarge has the same 2 NeuronCores as the inf2.8xlarge, just 16 GB host RAM instead of 128. The slim server (optb_server_slim.py) fits that by loading the host embedding model in bf16 and dropping the transformer layers (they run on-device), so the host footprint is a few GB:

docker run -d --device /dev/neuron0 --ipc=host -p 8080:8080 \
  xbill9/gemma4-optb-e4b:slim-devprefill

Validated on an inf2.xlarge: coherent output, device prefill ~0.11 s, ~36 tok/s, ~8 GB host RAM. (Compile still needs an inf2.8xlarge; the slim image reuses the same compiled model.)

The one non-obvious fix that made device prefill correct: Gemma-4's per-layer layer_scalar is a buffer, and NxD's ModelBuilder weight-sharding loads parameters only — so it must be copied from the checkpoint by hand, else every layer over-scales ~16× into garbage.

Why single-core doesn't work (and this does)

A single Option-B neff for E4B materializes ~15.4 GB of fp32 model constants, and a second neff on the same core tips it past the 16 GB NeuronCore budget → NRT_RESOURCE / status=4 Allocation Failure. bf16 shrinks the on-disk neff but not the on-device constants, so it doesn't rescue you. The only way to fit E4B is to shard the model across both cores (TP=2), so each rank holds ~half the weights (~7.7 GB/core).

Tensor-parallel then forces three problems that don't exist single-core. This build (tp_alias) solves all three:

1. 1.Co-residency. Under TP a neff spans both cores, so you can't park prefill on core 0 and decode on core 1. Fix: only the decode neff is resident on-device; prefill runs on the host CPU as a one-time seed. One neff resident → fits.
2. 2.Per-rank KV. Each rank computes a different KV head, but NxD's forward() returns only rank 0. Fix: KV is kept device-resident via input_output_aliases (never round-tripped through the host), and the initial seed is sliced head-r → rank r.
3. 3.GQA head-sharding. E4B has 8 query / 2 KV heads. Naïvely halving num_key_value_groups scrambles which query attends which KV head (the "4-vs-8" repeat_kv mismatch). Fix: column-shard k/v to 1 head per rank and keep `num_key_value_groups` unchanged.

Architecture

• —Sharding. q_proj/o_proj/k_proj/v_proj/gate/up/down are Column/Row-parallel across TP=2 (NxD ColumnParallelLinear/RowParallelLinear); k/v shard to nkv//TP = 1 head/rank.
• —Decode neff. Single-token forward against a fixed KV_MAX buffer. The K/V buffers are device-resident `nn.Parameter`s aliased as graph I/O (input_output_aliases), updated in place each step via a one-hot masked write (buf*(1-oh)+k*oh, trace-safe arithmetic).
• —Prefill (host, CPU). The fp32 reference model runs the prompt once on the host, producing the full KV cache + first token; the cache is sliced per-rank and copied into the decode neff's device parameters. This costs a one-time CPU pass per request (see Limitations).
• —Recipe basics (shared with the E2B Option-B port): eager attention, host-side PLE (Per-Layer Embedding) lookup, plain DynamicCache at trace time, explicit tanh-GELU + attention softcap, language model + head as real submodules.

Files

The neffs embed the base weights in bf16, so they are Apache-2.0 derivatives of google/gemma-4-E4B-it — see License below.

Run from these files

Requirements: an AWS `inf2.8xlarge` (2 NeuronCores) with the Neuron runtime, plus:

transformers==5.13.0
torch-neuronx==2.8.0  (SDK 2.23, neuronx-cc 2.23.6484.0)
neuronx-distributed>=0.17

# unpack a decode neff
tar xzf tpa_dec_512.tar.gz -C /workspace     # -> /workspace/tpa_dec

# validate + measure (CPU-seed prefill + device decode)
export KV_MAX=512 KV_BUCKET=128 MAXNEW=30
python tp_alias_run.py
# -> DEV GEN: 'The capital of France is **Paris**.'
#    SEQ_MATCH True
#    TP+ALIAS DECODE tok/s: 33.4 | 30 ms/tok

Recompile for a different context window (sizes are baked into the neff):

KV_MAX=2048 KV_BUCKET=512 python tp_alias_trace.py   # -> /workspace/tpa_dec

Compilation runs neuronx-cc across both ranks concurrently and peaks past host RAM on a 128 GB box — add swap before compiling (a 55 GB swapfile is enough); the resulting neff runs fine without swap.

Benchmarks (measured, inf2.8xlarge, us-west-2)

Per-token cost is essentially flat across context sizes because the KV cache is device-resident (no host round-trip). Correctness is exact: device greedy argmax matches the CPU float reference token-for-token, e.g. "The capital of France is Paris."

Limitations

• —`inf2.8xlarge` only — needs 2 NeuronCores (TP=2). Does not run on inf2.xlarge.
• —Prefill: the tp_alias build seeds on the host CPU (~1.4–1.6 s first token). The `tp_mb` device-prefill build above removes this (~0.16 s first token, prefill on-device via weight-sharing buckets) and is the recommended path for a low-latency endpoint.
• —Batch size 1, single-stream greedy/sampled decode. No continuous batching / paged attention.
• —Context is fixed at compile time; longer needs a recompile (compile needs swap).
• —Compiled specifically for Inferentia2 + the Neuron SDK versions pinned above.

License & attribution

• —Base model: google/gemma-4-E4B-it, © Google, Apache-2.0.
• —These artifacts (compiled neffs + scripts) are Apache-2.0 derivatives; the neffs contain the base weights in bf16. Redistributed under Apache-2.0 with attribution to Google — include the upstream LICENSE/NOTICE when redistributing.
• —The TP=2 tp_alias recipe (on-device aliased KV, CPU-seed prefill, GQA head-sharding) is released under Apache-2.0.

Not affiliated with or endorsed by Google or AWS. Provided for research/testing.