Add Aria text decoder (MoE) contrib model port

contrib/models/Aria/README.md

@@ -0,0 +1,80 @@
+# Contrib Model: Aria (Text Decoder)
+
+NeuronX Distributed Inference implementation of the Aria text decoder model.
+
+## Model Information
+
+- **HuggingFace ID:** `rhymes-ai/Aria`
+- **Architecture:** MoE Decoder-only transformer (LLaMA-based)
+- **Parameters:** ~3.9B total (text decoder)
+- **MoE Configuration:** 64 routed experts, top-6 routing, 2 shared experts
+- **Attention:** 20 Q heads, 20 KV heads (MHA), RoPE
+- **Layers:** 28 decoder layers
+- **License:** Check HuggingFace model card
+
+## Implementation Notes
+
+This port implements only the text decoder component of Aria.
+The vision tower and multi-modal projector are not included.
+
+The implementation reuses NXDI's standard MoE infrastructure (`initialize_moe_module`
+from `moe_v2`) which handles both routed and shared experts, following the same
+pattern as built-in Mixtral and Qwen3-MoE models.
+
+## Validation Results
+
+**Validated:** 2026-03-17
+
+| Metric | Value |
+|--------|-------|
+| Teacher-forced match | **98.12%** (PASSED) |
+| Greedy match | 72.81% |
+| Throughput (TKG) | 4.2 tokens/sec |
+| Config | tp=8, batch=1, seq_len=512, bf16 |
+| Instance | trn1.32xlarge |
+| HF golden reference | `AriaForConditionalGeneration` (text-only input) |
+
+### Token Match Details (10 prompts, 32 tokens each)
+
+| Prompt | Greedy | Teacher-Forced |
+|--------|--------|----------------|
+| "The theory of general relativity..." | 90.6% | 96.9% |
+| "The French Revolution began in..." | 100.0% | 100.0% |
+| "To solve a quadratic equation..." | 100.0% | 100.0% |
+| "Once upon a time in a distant galaxy..." | 12.5% | 96.9% |
+| "def fibonacci(n):..." | 100.0% | 100.0% |
+| "The Amazon River flows through..." | 96.9% | 93.8% |
+| "The concept of free will..." | 28.1% | 96.9% |
+| "To make a cup of coffee, first..." | 0.0% | 96.9% |
+| "List three benefits of regular exercise..." | 100.0% | 100.0% |
+| "If all roses are flowers..." | 100.0% | 100.0% |
+
+Note: HF golden uses full `AriaForConditionalGeneration` with text-only input
+(no pixel_values). `AriaTextForCausalLM` cannot be loaded directly due to
+weight prefix mismatch (`language_model.*` in checkpoint). Greedy divergence
+is expected for MoE models in bf16 due to cascading expert routing differences.
+
+## Compatibility Matrix
+
+| Instance/Version | 2.20+ | 2.19 and earlier |
+|------------------|-------|------------------|
+| Trn1             | Working | Not tested |
+| Inf2             | Not tested | Not tested |
+
+## Performance
+
+Profiled on trn1.32xlarge (single NeuronCore utilization):
+
+| Metric | Context Encoding | Token Generation |
+|--------|-----------------|------------------|
+| Throughput | - | 4.2 tok/s |
+| MBU (Memory) | 1.4% | 2.8% |
+| MFU (Compute) | 0.9% | 0.0% |
+
+*Batch size 1, sequence length 512, BF16 precision, TP=8*
+
+## Maintainer
+
+Neuroboros Team - Annapurna Labs
+
+**Last Updated:** 2026-03-17

contrib/models/Aria/src/__init__.py

@@ -0,0 +1 @@
+from .modeling_aria_text import NeuronAriaTextForCausalLM, AriaTextInferenceConfig