Zek-Takai/glm53-flash-internals

GLM-5.3-Flash Internals — a measured profile

A measured characterization of the internals of zai-org/GLM-5.3-Flash (glm5_next) at real serving scale: activation landscapes, attention-sink anatomy, FP8 working margins to 260k context, sparse-indexer selection maps, and the memory behavior of a two-node vLLM serve. Every finding cites its capture and states its conditions. profile.json beside this card is the machine-readable anatomy.

Produced with Claude Code and Spectra Scope, a model-anatomy inspector that hooks a serving model's modules and records what actually flows through them.

Companion dataset: glm53-flash-harvest — 117.4M on-policy tokens generated by the same model, for drafter/speculator training.

⚠️ Precision context — read before using any number

Every measurement here was taken on a mixed-precision quant, not the BF16 release. The checkpoint under test ("R1", 179.85 GiB, built with ModelOpt MIXED_PRECISION on top of the LibertAIDAI NVFP4 weight-only release):

• routed experts: NVFP4 (E2M1, FP8-E4M3 per-16 block scales)
• 138 modules FP8 per-channel-per-token: shared-expert MLPs and dense-layer MLPs (including the draft-namespace duplicates for the MTP layer)
• BF16: all attention (KDA + MLA + indexer), embeddings, lm_head, the vision tower, MTP-connective tensors, all norms

Consequences for reading the findings: attention-side amplitudes are unquantized truth; MLP-side sites are read at the FP8/FP4 boundary, so QuantFP8-site figures are what the quant op actually ingests. A BF16 serve may behave differently, particularly on MLP-side amplitudes and anywhere quant boundaries clip or rescale; the structural findings (sink anatomy, family temperaments, selection behavior) are expected to transfer, but treat absolute magnitudes as quant-specific until reproduced on other precisions.

Serve rig for all captures: 2× NVIDIA DGX Spark (GB10, sm_121), vLLM built from PR #53906, TP=2 over a 200G rail, --enforce-eager (CUDA graphs and compile swallow forward hooks; production serving uses graphs and is faster — eager is the capture condition), text-only, MTP k=5 speculative decoding ON, max context 262,144, KV byte-capped, Marlin MoE backend.

Instrument: spectrascope-companion (0.1.0 → 0.2.2 across the batteries) — torch global forward hooks, per-module input aggregates (amax / absmean / element count) accumulated on-device while armed, drained per TP worker. Limits at capture time are stated inline per battery.

Anatomy (from the checkpoint — facts, not guesses)

Decoder: 45 layers, hidden 4096, vocab 154,880, native context 1,048,576 (served here at 262,144). Two interleaved attention families:

• 34 KDA linear-attention layers (indices 0–2, then 3-of-4 up to 44): 64 heads × 128, short-conv kernel 4, learned decay gates (lower bound −5.0). Recurrent constant-size state per layer — no KV growth. These layers are why host memory stays flat and the long-context tax is small.
• 11 DeepSeek-sparse-attention (MLA) layers (3, 7, 11, … 43 — every 4th): 64 heads, q_lora 1536 / kv_lora 512, qk_head 256, v_head 256, all-NoPE (qk_rope_head_dim = 0 — no rotary anywhere). Each carries a sparse indexer: 32 index heads × 128 select top-2048 tokens per query from a 4-deep compressed key pool, tail always selected, and the indexer is shared with the MTP iteration.

MoE: first 3 layers dense (intermediate 12,288), the other 42 sparse: 288 routed experts + 1 shared expert, 8 active per token, expert intermediate 2048, sigmoid scoring with noaux_tc top-k, routed scaling 2.5, router computed in fp32.

mHC hyper-connections (mhc: true, mult 4): the residual stream is 4-way multiplexed with per-layer mixing weights normalized by 20 sinkhorn iterations. "The residual stream" is plural in this model — any residual-based analysis must account for 4 streams and their mixing.

MTP: one in-checkpoint next-token-prediction layer (index 45) — a speculative drafter. The draft shares the sparse indexer.

Vision tower: 24-layer ViT (1024 hidden, patch 14, merge 2 → 4096 out), BF16 untouched by every quant. Text-only serves keep it unexercised.

Open questions this profile exists to answer

1. Where does position live? With zero rotary, position information must ride the KDA recurrences and conv kernels.
2. What does the sparse indexer actually select? Top-2048 of 262k is <1% — locality vs. semantic retrieval, tail behavior, per-head specialization, per-depth shifts.
3. Expert routing shape at 288/8 with sigmoid-noaux: load balance, per-domain specialization, shared-expert load vs. routed.
4. How do the 4 hyper-connection streams divide labor?
5. Activation magnitude landscape per family: KDA in-projections vs MLA projections vs expert inputs vs MTP.
6. MTP acceptance anatomy: where the drafter fails (acceptance ~3.9–4.1 of 5 at temp 0.7/greedy).

Findings

Append-only; each battery names its captures. Amplitudes are input amax unless stated. "TP0/TP1" are the two tensor-parallel workers.

First live captures (2026-08-28, companion 0.1.0)

Captures bridge-20260828-063027 (120-token prompt), wild-aurora (400-token prompt). 102 modules captured across a ~180G two-node serve.

• Massive activations confirmed on this family: RMSNorm inputs peak at amax ~6016 on both ranks — the same massive-activation channel class documented on other model families, now on glm5_next.
• mHC is hook-visible: mhc_post_op fires as a module boundary (amax ~900) — the hyper-connection lane is capturable without engine surgery.
• Site magnitude hierarchy (51 named sites/rank): RMSNorm ~6016 ≫ mhc_post_op ~900 ≫ decoder-layer input ~466 ≫ shared_experts.down_proj 68.5/41.5 (sharded; ranks diverge) ≫ QuantFP8 op 46.5 ≫ MLA attention core 22.4 ≫ KDA gate_proj 13.9 ≫ router 7.1 ≫ dense mlp.down_proj 0.6. Norm-dominated: the classic massive-activation signature.
• Replicated vs sharded tensors identified by rank symmetry (matching amax = replicated; diverging = sharded) — measured, not inferred from code.
• Instrument limit at this stage: decoder names collapsed across layers (site-level truth only; fixed by battery 5).

Battery 1 — domain & regime separation (2026-08-28)

Captures recall-baseline, math-reasoning, code-gen, long-context (~7k needle doc), prefill-only (same doc, 1 token out), decode-heavy.

1. RMSNorm peak sits at ~6000 (5984–6016) in every run with real decode, constant across domains — a structural attractor, not input-driven. A 7,000-token prefill with 1 generated token peaked at just 69. [Superseded by battery 2, finding 1: the split is position-anchored, not decode-anchored — the low prefill peak was a prefix-cache artifact.]
2. FP8 trunk working range: QuantFP8-site input amax grows with task/context: 30 (recall) → 50 (math) → 52 (code) → 68 (7k ctx). FP8-E4M3 saturates ~448 → ~6.6× headroom at 7k.
3. Family temperament: KDA sites hold a tight band in every condition (g_b_proj 13.4±0.5) — constant-state temperament. MLA core input scales with context (10.8 → 21.3) — reach-driven.
4. MTP always-on (hypothesis falsified honestly): all 23 MTP site entries fire even in prefill-only — the drafter engages from the first generated token, just cooler (eh_proj 2.31 vs ~3.15).
5. Rank-symmetry catalog: 39 sites identical on both ranks in every condition (replicated: norms, routers, in-projs, indexer, mHC); 12 diverge (sharded: down_proj, o_proj, o_norm, act_fn, MLA core, QuantFP8) — a measured TP ownership map.
6. Sparse retrieval works: a 7k needle answered correctly through the top-2048 indexer.

Battery 2 — context ladder 4k–100k + unique-text controls (2026-08-28)

Captures ladder-4k/16k/32k/64k/100k (needle at ~55% depth), controls ctl-4k / ctl-32k (fully unique text, no shared prefix).

Ladder amax (TP0): RMSNorm 5984 / 5984 / 79 / 95 / 97.5; mhc_post_op 940 / 940 / 117 / 130 / 156; decoder-layer 440 / 440 / 19.5 / 18.1 / 17.8; QuantFP8 67.5 / 65.5 / 59.8 / 55.5 / 59.8; MLA core ~20.8 at every rung; KDA gates 11.8–13.9 flat. Unique-text ctl-32k restored RMSNorm 5984 / mhc 940 / layer-in 440 with 7.5× the element count.

1. The massive activation is POSITION-anchored (correction to battery 1): the ~6000 attractor belongs to the early-sequence positions (BOS/sequence-start neighborhood). Any run computing the sequence start fresh shows it at any size or domain; any run whose opening positions come from prefix cache does not. Late-position forwards peak ~18–100. Classic attention-sink / massive-activation anatomy, measured at serving scale, constant at ~5984–6016 regardless of content.
2. Prefix caching is live on this serve branch (contrary to early ecosystem notes). Protocol consequence: activation-capture prompts must be prefix-unique (or the cache flushed), or early-position physics silently drop out of the landscape.
3. FP8 margin: safe through 100k. QuantFP8 amax never exceeded ~68 on the ladder (fresh controls 46–54.5); no climb toward the 448 ceiling.
4. MLA amplitude saturates by design: the battery-1 rise flattens at ~20.8 for 16k–100k — the top-2048 selection cap bounds what attention ingests once context ≫ topk. The sparse design shows up in the physics.
5. Retrieval: 5/5 needles correct 4k→100k plus a unique-doc control.
6. KDA gates, router, indexer input: flat at all depths.

Battery 3 — monitored memory rerun (2026-08-28)

Captures ladder2-4k…100k, all prefix-unique; host and per-process RSS sampled through the run with a settle window.

1. No leak — a bounded one-time commit. Head-node memory 110.03G → peak 110.88G → settled 110.52G: a full ladder including a fresh 100k prefill added zero net retained memory. The earlier ~1.5G rise across batteries 1–2 was first-touch page commit of the pre-granted KV/prefix pool and eager workspaces — bounded by the KV grant, not open-ended.
2. Attribution clean: capture engine RSS flat (0.58G), serve tree host RSS flat (4.14G head / 1.90G worker) throughout — the transient lived in the CUDA/unified-memory pool, not in any process heap.
3. Fresh-100k control: RMSNorm 5984 / mhc 940 (attractor holds at depth, fully unique text); QuantFP8 58.5; MLA 18.5.
4. Honest fresh-prefill timing (eager capture rig, end-to-end with ~200-token answers): 4k 9s / 16k 23s / 32k 35s / 64k 63s / 100k 96s — ~1.2–1.3k tok/s aggregate prefill. (Battery 2's faster big-rung timings were prefix-cache-flattered.)
5. Retrieval: 5/5 unique-doc needles.

Battery 4 — to the serving ceiling (2026-08-28)

Captures ladder2-160k, ladder2-250k-b (= 260,408 tokens exact, counted by the serve's own token counter), retention-260k-repeat, rim-test-0..3.

1. Physics hold at the ceiling: attractor 5984 / mhc 940 at 160k and 260k fresh; QuantFP8 54.75–59.5 — final FP8 verdict: ≥8× margin at maximum servable context, the quantized trunk never approaches the E4M3 ceiling at any depth; MLA saturated ~19.4–19.6; KDA flat.
2. Retrieval at the ceiling: correct. Cumulative 8/8 fresh-text needles 4k→260k through the top-2048 selector.
3. Memory model validated both directions: 160k retained nothing; the first-ever 260k run retained +0.74G/+0.94G (head/worker) — sized like the KV footprint of the newly-touched depth. The falsifiable prediction (a repeat 260k retains ~zero) was then confirmed: −0.00G / −0.02G, and the repeat ran in 13s vs 207s (prefix cache served the identical document wholesale).
4. Rim test — leak hypothesis rejected: four fresh unique documents (~160k new tokens) forced wholesale KV-block eviction through a fully cycled pool (cumulative unique tokens ~636k > pool capacity ~592k): head −0.45G, worker +0.13G, all needles correct through the churn. The observed "staircase" was the pre-granted KV pool committing physical pages on first touch until full — a rim, not a leak. Freed blocks return to the pool's free list; a living allocation's pages return only at serve restart. Standing signal: steady-state usage marching meaningfully past the rim would indicate a real leak.
5. Timing: 160k end-to-end 153s; 260k 207s — ~1.3k tok/s aggregate eager prefill, remarkably flat across the whole ladder.
6. Operational note: prompts over max_model_len are rejected (word-count estimates drifted ~13% low on this text; the serve's count endpoint is the honest sizing tool but errors above the limit itself).

Battery 5 — the sink hunt: first per-layer captures (2026-08-28)

Captures shakedown-020, sink-hunt (4k fresh), selection-map (32k fresh); companion 0.2.0 with per-layer naming and worker ranks — 1,476 modules per capture (738 per TP worker).

1. The massive activation localized: final-norm / stream-collapse anatomy, not any decoder layer. Inside all 45 layers the hottest submodule stays ≤ ~100. The chain that builds the ~6000: per-layer residual (hyper-connection stream bundle) runs spiky up to ~456 (hot layers #18 456, #22 448, #34 376, #43–44 378–440 — non-monotonic: different layers read different-magnitude stream mixes) → the model's only explicit mhc_post_op sits at layer 44 and collapses the 4 streams (input 940) → the final model RMSNorm ingests 5984. Combined with battery 2: the sequence-start token's residual grows through depth and crystallizes at stream-collapse + final norm — textbook attention-sink anatomy, measured module-by-module at serving scale.
2. mHC architecture note (measured): stream expansion/mixing is implicit per layer; a single explicit post-op at layer 44 performs the collapse. The mHC figure earlier batteries tracked was layer 44's collapse point all along.
3. Per-layer temperaments: KDA gate inputs grow gently with depth (5 → 15); shared-expert down_proj is the within-layer hot spot, spiking at layers 3 and 7 (exactly 100.0 — a suspiciously round ceiling, possibly activation-clamp-related; open question) and running 40–75 in late MoE layers.
4. Instrument: the selection lane came up empty this boot — root-caused live (on this vLLM branch the indexer's topk_indices_buffer attr is None; indices ride the module's integer return, which a float-only output tap skipped). Fixed in companion 0.2.1.

Battery 6 — first selection maps (2026-08-28)

Captures selection-map-c (32k), selection-deep-c (100k); companion 0.2.2 (three real serving-stack traps found and fixed against the living serve: None buffer attr → integer-return fallback → −1-padding rows; each fix regression-tested).

1. First indexer selection maps at serving scale. 128-bucket position histograms of what the top-2048 selector actually read: at 32k, a structured non-uniform coverage profile across the whole document; at 100k, selection mass shifts strongly toward later regions with a bright final bucket and a distinct mid-document band near the needle's 55% position. The MTP indexer's map differs from the main layers' (sparser, tail-heavier) — the drafter reads the sequence differently than the verifier.
2. Instrument discovery — the layers alias one buffer: all 11 DSA layers reported bit-identical maps; the indexer op writes a shared model-level indices tensor, so per-module references collapse to last-writer-wins. Per-layer selection differentiation needs clone-on-stash (bounded copy at forward time) — planned; until then each capture carries one authoritative main-path map (last writing layer, last chunk) plus the MTP map.
3. Calibration notes: counts (~4.19M = 2048 queries × 2048 selections) show the stash holds the final prefill chunk's selection rather than the last decode step — decode-path stashing to be verified after the de-alias fix. Tail-share reads near zero at 32k despite the always-select-tail law — the compressed-pool tail may ride a separate mechanism; open question, not overclaimed.

Known limits & open items

• Per-layer selection maps blocked by the shared-buffer aliasing (fix planned); decode-path selection stashing unverified.
• Vision tower unexercised (text-only serves); multimodal behavior of the quant untested.
• Outputs, sequences, and attention probabilities not recorded by the instrument at these versions — input aggregates only.
• Expert-routing statistics (open question 3) and hyper-connection stream labor division (question 4) not yet captured.
• All timings are from the eager capture rig; production (CUDA-graphs) serving is materially faster.

Provenance & citation

Measurements taken 2026-08-28 with Spectra Scope and Claude Code against a vLLM PR-#53906 serve of the R1 mixed-precision requant described above. Base model: zai-org/GLM-5.3-Flash (MIT).

@misc{glm53flashinternals2026,
  title  = {GLM-5.3-Flash Internals: a measured profile},
  author = {Zek-Takai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Zek-Takai/glm53-flash-internals}
}