<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>LLM Architecture Knowledge Base</title>
    <link>https://lizeman.github.io/llm-arch-kb/</link>
    <description>A chronological reference of decoder-only LLM architecture innovations — math, simple implementations, tradeoffs, and adopting models with citations.</description>
    <language>en</language>
    <atom:link href="https://lizeman.github.io/llm-arch-kb/rss.xml" rel="self" type="application/rss+xml" />
    <item>
      <title><![CDATA[DeepSeek Sparse Attention (DSA)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/dsa/</link>
      <description><![CDATA[A two-stage attention: a fast Lightning Indexer scores how relevant each historical key is to the current query, then full attention runs only over the top-K keys. Built on top of MLA — the latent KV cache is unchanged; the sparsity is in which cached entries get attended.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/dsa/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[Kimi K2 MoE (K2 MoE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/kimi-k2-moe/</link>
      <description><![CDATA[Trillion-parameter MoE with 384 routed experts plus one shared expert — wider than DeepSeek V3's 256 + 1 by ~50%. Stabilized at scale by the MuonClip optimizer (Muon family with weight clipping). Same shared-expert + aux-loss-free routing lineage as DeepSeekMoE.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/kimi-k2-moe/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Gemma 3 Norm-Everywhere (Norm-Everywhere)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/gemma3-norm-everywhere/</link>
      <description><![CDATA[Sandwich-LN applied to BOTH the attention and FFN sublayers — four RMSNorms per block. Combined with QK-Norm and interleaved local/global attention, it stabilizes the entire Gemma 3 size family from a single recipe.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/gemma3-norm-everywhere/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[OLMo 2 Reordered Post-Norm (OLMo 2 Post-Norm)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/olmo2-post-norm/</link>
      <description><![CDATA[Move both RMSNorms to sit after each sublayer output, inside the residual branch — recovering Post-Norm's bounded residual stream while keeping Pre-Norm's gradient flow. Trained stably at 13B by combining with QK-Norm and an output-side Z-loss.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/olmo2-post-norm/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[nGPT — Normalized Transformer on the Hypersphere (nGPT)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/ngpt/</link>
      <description><![CDATA[Project every vector to unit norm at every step. Eliminates LayerNorm/RMSNorm by structural construction; cosine similarity replaces inner product everywhere. Trains 4-20× faster than baseline GPT at matched quality, per the paper.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/ngpt/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[Hyper-Connections (HC)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/residual/hyper-connections/</link>
      <description><![CDATA[Generalize the residual connection to n parallel streams. Each sublayer learns depth-wise read/write coefficients that mix streams in and out. Recovers the standard residual at n=1 and gives a strict superset of expressivity above that.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/residual/hyper-connections/</guid>
      <category>residual</category>
    </item>
    <item>
      <title><![CDATA[Dynamic Hyper-Connections (DHC / mHC)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/residual/dynamic-hc/</link>
      <description><![CDATA[The per-token variant of Hyper-Connections: A_l and B_l are computed as small projections of the current input rather than learned constants. Each token gets its own routing pattern across the n streams, at the cost of one extra projection per layer.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/residual/dynamic-hc/</guid>
      <category>residual</category>
    </item>
    <item>
      <title><![CDATA[Auxiliary-Loss-Free Load Balancing (Aux-Loss-Free)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/aux-loss-free/</link>
      <description><![CDATA[Maintain a per-expert bias term that shifts the gate's top-K decision toward under-used experts. The bias enters at selection time only — the final expert outputs are weighted by the original (unbiased) gate values. Load balance without quality cost.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/aux-loss-free/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Multi-Head Latent Attention (MLA)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/mla/</link>
      <description><![CDATA[Compress K and V to a small per-token latent; reconstruct heads at attention time. ~5–7× smaller KV cache than MHA on DeepSeek-V2 ablations.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/mla/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[Decoupled RoPE (Decoupled RoPE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/decoupled-rope/</link>
      <description><![CDATA[Split the K projection into a small position-aware head that carries RoPE and a larger position-free body that absorbs into the latent KV cache. The trick that lets MLA compress K, V without losing position.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/decoupled-rope/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[LongRoPE — Per-Dimension RoPE Search (LongRoPE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/longrope/</link>
      <description><![CDATA[Use evolutionary search to find per-dimension RoPE rescaling factors. Generalizes YaRN's closed-form frequency-band recipe to arbitrary non-monotone schedules. Demonstrated 2M+ context extension on Llama-2 with a short fine-tune.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/longrope/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[Mixtral-Style Coarse MoE (Mixtral MoE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/mixtral-moe/</link>
      <description><![CDATA[Eight full-sized SwiGLU experts per MoE layer; top-2 routing with a standard load-balancing auxiliary loss. The Mistral-org variant that brought open-weights MoE into wide use, sitting on the coarse end of the granularity spectrum that DeepSeekMoE later contested.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/mixtral-moe/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Lightning Attention (Lightning)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/lightning-attention/</link>
      <description><![CDATA[Tile and fuse linear attention's prefix-sum recurrence so it runs faster than FlashAttention at long context. Interleaved 7:1 with softmax attention in MiniMax-01 to recover what linear attention loses on absolute quality while keeping its linear-in-T scaling.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/lightning-attention/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[DeepSeekMoE (DeepSeekMoE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/deepseek-moe/</link>
      <description><![CDATA[Split the FFN into many small experts, route each token to top-K, and reserve a few always-on shared experts for common-knowledge work. Activated parameter count stays small; specialization sharpens.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/deepseek-moe/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Activation Beacon (Activation Beacon)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/long-context/activation-beacon/</link>
      <description><![CDATA[Insert learned 'beacon' tokens that summarize past activation windows. The model attends to beacons instead of the full long-context activations; 100× context extension with a tiny fine-tune of just the beacon parameters.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/long-context/activation-beacon/</guid>
      <category>long-context</category>
    </item>
    <item>
      <title><![CDATA[StreamingLLM and Attention Sinks (Attention Sinks)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/long-context/streaming-llm/</link>
      <description><![CDATA[The first 1-4 tokens of any pretrained decoder act as attention sinks — they absorb the softmax mass that has nowhere else to go. Pin them in the KV cache and you can slide the rest of the window over arbitrarily long input without quality collapse.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/long-context/streaming-llm/</guid>
      <category>long-context</category>
    </item>
    <item>
      <title><![CDATA[YaRN — Yet Another RoPE eXtensioN (YaRN)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/yarn/</link>
      <description><![CDATA[Treat RoPE's rotation bands as three regimes — preserve the fast ones, linearly interpolate the slow ones — and rescale the softmax temperature. The 2023 long-context extension recipe of choice.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/yarn/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[NTK-Aware RoPE Scaling (NTK-Aware)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/ntk-aware-rope/</link>
      <description><![CDATA[Multiply RoPE's base b by s^(d_h/(d_h-2)) for extension factor s. Fast dimensions (low index) are nearly untouched; slow dimensions get linearly interpolated. Zero fine-tuning needed at small extensions; the inspiration for YaRN's per-band approach.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/ntk-aware-rope/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[LongNet — Dilated Attention (LongNet)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/long-context/longnet/</link>
      <description><![CDATA[Each attention head operates at a different dilation rate r ∈ {1, 2, 4, ...}: stride-r local windows. Combining log₂(N) such heads gives O(N · log N) total cost with full reachability across the sequence. Demonstrated training on 1B-token sequences.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/long-context/longnet/</guid>
      <category>long-context</category>
    </item>
    <item>
      <title><![CDATA[Position Interpolation (PI)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/position-interpolation/</link>
      <description><![CDATA[Divide position values by the extension factor s before applying RoPE. Position t becomes t/s; trained rotation angles never extrapolate. Simple, parameter-free, works — but loses resolution uniformly across all frequency bands.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/position-interpolation/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[Landmark Attention (Landmark)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/long-context/landmark-attention/</link>
      <description><![CDATA[Insert a learnable landmark token at the end of each chunk; train it to summarize the chunk's content. At inference, queries attend to landmarks to gate which chunks get full attention. Hierarchical chunked attention with O(C + K·W) cost where C is the chunk count, K the chunks selected, and W per-chunk size.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/long-context/landmark-attention/</guid>
      <category>long-context</category>
    </item>
    <item>
      <title><![CDATA[Grouped-Query Attention (GQA)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/gqa/</link>
      <description><![CDATA[Group query heads so each group reads one shared K, V pair. 4-8× KV-cache reduction with quality near MHA; the dominant attention layout for dense LLMs from 2023 onward.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/gqa/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[FlashAttention (FlashAttention)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/flash-attention/</link>
      <description><![CDATA[Tile attention so QK^T and the softmax stay in SRAM rather than round-tripping through HBM. Exact (not approximate) attention, 2–4× faster, 5–20× less peak memory. The universal kernel under every modern transformer trainer.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/flash-attention/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[No Position Encoding (NoPE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/nope/</link>
      <description><![CDATA[Drop the positional encoding entirely. The causal mask asymmetrically restricts each query to its left context; the model learns to extract token position from that asymmetry alone. Works surprisingly well; reveals what positional encodings are actually adding.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/nope/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[Memorizing Transformers (Memorizing)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/long-context/memorizing-transformers/</link>
      <description><![CDATA[Augment one mid-stack attention layer with a kNN lookup over a non-differentiable bank of past K, V pairs. Retrieved keys/values are merged into the layer's attention. Conceptually adjacent to RAG; mechanically a forerunner.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/long-context/memorizing-transformers/</guid>
      <category>long-context</category>
    </item>
    <item>
      <title><![CDATA[DeepNet — Scaling Transformers to 1000 Layers (DeepNet)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/residual/deepnet/</link>
      <description><![CDATA[Scale each Post-Norm residual addition by a constant α(N) and shrink sublayer-weight init by β(N), both derived from depth. Trains 1000-layer encoder-decoders without divergence; gives Post-Norm geometry back to deep transformers that had been pushed toward Pre-Norm purely for stability reasons.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/residual/deepnet/</guid>
      <category>residual</category>
    </item>
    <item>
      <title><![CDATA[NormFormer — Extra Normalization in the Residual (NormFormer)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/residual/normformer/</link>
      <description><![CDATA[Three extra normalizations bolted onto Pre-Norm: a LayerNorm on the attention output, learnable per-head scaling of that output, and a LayerNorm inside the FFN after the first linear. 24% faster pretraining at matched final perplexity on the 1.3B baseline; not adopted in production but cleanly motivated.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/residual/normformer/</guid>
      <category>residual</category>
    </item>
    <item>
      <title><![CDATA[Attention with Linear Biases (ALiBi)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/alibi/</link>
      <description><![CDATA[Bias attention scores by a per-head linear function of the query-key distance. No learnable position parameters; extrapolates beyond training length without any fine-tune. Lost the dominance race to RoPE for dense decoders but is mechanically illuminating.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/alibi/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[Sandwich-LN (Sandwich-LN)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/sandwich-ln/</link>
      <description><![CDATA[Two RMSNorms per sublayer: one before, one after, both inside the residual branch. Bounds the per-block contribution to the residual stream while keeping the identity path unnormalized. Used by Gemma 2 and 3.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/sandwich-ln/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[Rotary Position Embedding (RoPE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/rope/</link>
      <description><![CDATA[Rotate the (2i, 2i+1) coordinate pairs of Q and K by an angle proportional to position. Relative position falls out of the inner product for free; the model never sees a learned position table.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/rope/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[Switch Transformer (Switch)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/switch-transformer/</link>
      <description><![CDATA[Top-1 expert routing — each token goes to exactly one expert. Simpler than GShard's top-2 baseline. Trillion-parameter Switch-C model proved sparse MoE was tractable at unprecedented scale; established the load-balancing loss formulation everyone uses.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/switch-transformer/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Query-Key Normalization (QK-Norm)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/qk-norm/</link>
      <description><![CDATA[Normalize Q and K before the attention dot product so the logit magnitude stays bounded — a near-free fix for the outlier blow-ups that wreck attention at long context.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/qk-norm/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[Performer — Random Feature Softmax Approximation (Performer)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/performer/</link>
      <description><![CDATA[Approximate the softmax kernel as ⟨φ(q), φ(k)⟩ for a random feature map φ. The decomposition reorders QKV matmuls so attention is O(N·d) per layer; unlike Linformer the approximation is sequence-length-independent and works for autoregressive decoders.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/performer/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[BigBird (BigBird)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/bigbird/</link>
      <description><![CDATA[Each query attends to three sets of keys: a random sample, a local window, and a fixed global set. O(N) attention; provably universal-approximator and Turing-complete. The 2020 successor to Sparse Transformer that put sparse attention on a theoretical footing.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/bigbird/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[Linformer — Low-Rank Attention Projection (Linformer)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/linformer/</link>
      <description><![CDATA[Multiply K and V by learned projection matrices E, F ∈ ℝ^{k×N} that compress the sequence axis to a fixed rank k. Attention becomes O(N·k) per query. Empirically near-softmax on encoder tasks; the sequence-length-dependent projection makes decoder use awkward.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/linformer/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[Linear Attention (Linear Attention)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/linear-attention/</link>
      <description><![CDATA[Replace softmax(QK^T)V with φ(Q)·(φ(K)^T·V). The K^T·V term is a small (d×d) state independent of sequence length; the operator becomes a recurrence at inference. Approximate, but linear in T.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/linear-attention/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[GShard (GShard)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/gshard/</link>
      <description><![CDATA[Top-2 routing across hundreds of experts in a sharded transformer encoder. Established expert capacity, dispatch tensors, and the per-expert random-token-dropping policy. The conceptual ancestor of Switch, Mixtral, and DeepSeekMoE.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/gshard/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Sliding Window Attention (SWA)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/sliding-window/</link>
      <description><![CDATA[Restrict each query to the last W keys. Per-layer compute becomes O(W) per token; stacked layers still cover the full context because the receptive field grows linearly with depth.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/sliding-window/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[ReZero — Residual With Learnable Skip Scale (ReZero)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/residual/rezero/</link>
      <description><![CDATA[Scale each residual branch by a per-layer learnable scalar α initialized to zero. The network starts at identity (gradient = 1 through every block) and learns to integrate sublayer contributions gradually. Stable at depths where standard transformers diverge.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/residual/rezero/</guid>
      <category>residual</category>
    </item>
    <item>
      <title><![CDATA[Swish-Gated Linear Unit (SwiGLU)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/swiglu/</link>
      <description><![CDATA[Replace the FFN with a Swish-gated bilinear form: one linear path multiplied element-wise by Swish of another. Consistent quality win over ReLU/GELU at matched parameter count; the default modern FFN.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/swiglu/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[ReLU-Gated Linear Unit (ReGLU)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/reglu/</link>
      <description><![CDATA[Bilinear gated FFN with a ReLU-activated gating path. Matches SwiGLU and GeGLU on quality within noise per Shazeer's original ablations; the rarely-shipped third GLU variant.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/reglu/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Pre-Norm, Post-Norm, and Sandwich Placement (Norm Placement)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/norm-placement/</link>
      <description><![CDATA[Pre-Norm became the default because it lets gradients flow unaltered through the residual stream and removes the need for warmup; sandwich placements like Gemma 3's norm-everywhere are the modern minority alternative.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/norm-placement/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[GELU-Gated Linear Unit (GeGLU)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/geglu/</link>
      <description><![CDATA[Bilinear gated FFN with a GELU-activated gating path. Matches SwiGLU on quality; lineage difference is mostly historical, with the Gemma family carrying GeGLU forward and the Llama/DeepSeek family carrying SwiGLU.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/geglu/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Reformer — LSH Attention (Reformer)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/reformer/</link>
      <description><![CDATA[Hash Q and K to discrete buckets via random rotation; restrict attention to queries and keys in the same bucket. O(N · log N) compute via learned (not hand-designed) sparsity. Lost to FlashAttention's I/O-aware exact computation on practical grounds.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/reformer/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[Multi-Query Attention (MQA)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/mqa/</link>
      <description><![CDATA[One K, V projection shared by every query head. H× smaller KV cache than MHA; small but real quality drop that motivated GQA. The first move in the MHA → MQA → GQA → MLA evolution.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/mqa/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[Compressive Transformer (Compressive)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/long-context/compressive-transformer/</link>
      <description><![CDATA[Two-tier memory: a short-term FIFO cache plus a long-term compressed memory built by applying a learned compression function (mean pooling, conv, or attention) to evicted segments. The original long-context recipe that traded resolution for length.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/long-context/compressive-transformer/</guid>
      <category>long-context</category>
    </item>
    <item>
      <title><![CDATA[Root Mean Square Layer Normalization (RMSNorm)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/rmsnorm/</link>
      <description><![CDATA[Normalize by RMS instead of subtracting the mean — drop one statistic, one bias, and ~5-10% of normalization compute, with no measurable quality cost.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/rmsnorm/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[Sparse Transformer (Sparse Transformer)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/sparse-transformer/</link>
      <description><![CDATA[Factorize attention into two structured sparse patterns: strided (local window) and fixed (long-range hubs). Total cost O(T·√T) per layer; full reachability in two hops. Foundational for sparse-attention research; superseded by SWA, BigBird, MoE attention in production.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/sparse-transformer/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[Sinusoidal Position Encoding (Sinusoidal)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/positional/sinusoidal/</link>
      <description><![CDATA[Add a fixed sinusoid of geometrically-spaced frequencies to every token embedding. Closed-form, no learned parameters, extrapolates moderately. The 2017 transformer's original answer to 'how does the model know where each token is'.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/positional/sinusoidal/</guid>
      <category>positional</category>
    </item>
    <item>
      <title><![CDATA[Multi-Head Attention (MHA)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/attention/mha/</link>
      <description><![CDATA[The original attention layout: H independent heads, each computing scaled dot-product attention on a d_h-dim subspace, concatenated then projected. Now superseded by GQA, MLA in production decoders, but still the conceptual foundation.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/attention/mha/</guid>
      <category>attention</category>
    </item>
    <item>
      <title><![CDATA[FFN with ReLU (FFN-ReLU)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/ffn-relu/</link>
      <description><![CDATA[The original transformer FFN: ReLU squashing of a 4× hidden expansion, projected back. Foundational and historical — every later FFN variant (GELU, SwiGLU, GeGLU) is a refinement of this two-layer-with-nonlinearity pattern.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/ffn-relu/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Sparsely-Gated MoE (Sparse MoE)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/sparse-moe/</link>
      <description><![CDATA[Top-K gating with per-expert noise plus an importance-balancing loss. The 2017 paper that established that 'sparse activation + load balance' is a viable training-time recipe; every modern MoE descends from it.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/sparse-moe/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[Layer Normalization (LayerNorm)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/normalization/layernorm/</link>
      <description><![CDATA[Per-token feature-axis normalization with learnable scale and shift. The 2016 building block the original transformer used; superseded in modern decoders by RMSNorm, which drops the mean-centering step at no quality cost.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/normalization/layernorm/</guid>
      <category>normalization</category>
    </item>
    <item>
      <title><![CDATA[Gaussian Error Linear Unit (GELU)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/ffn-moe/gelu/</link>
      <description><![CDATA[Multiply x by the standard normal CDF: GELU(x) = x · Φ(x). Smoother than ReLU, weighted by the input's z-score. The default transformer FFN activation in BERT, GPT-2, GPT-3 era; superseded by gated variants (SwiGLU, GeGLU) but still the substrate they sit on.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/ffn-moe/gelu/</guid>
      <category>ffn-moe</category>
    </item>
    <item>
      <title><![CDATA[The Residual Stream (Residual)]]></title>
      <link>https://lizeman.github.io/llm-arch-kb/residual/residual-overview/</link>
      <description><![CDATA[Every modern transformer block adds its output to the running activation rather than replacing it. The 'residual stream' is the linear path through depth that carries information unaltered; each block contributes a perturbation.]]></description>
      <pubDate>Tue, 12 May 2026 00:00:00 GMT</pubDate>
      <guid isPermaLink="true">https://lizeman.github.io/llm-arch-kb/residual/residual-overview/</guid>
      <category>residual</category>
    </item>
  </channel>
</rss>