Repository navigation

[Perf] [MI355X] glm53: Shared layer and MoE hand-offs, NextN entry, K-split front, indexer launch, MoE expert dedup and weight loads, kernarg reload, LM-head loads - #85

Merged

jhinpan merged 25 commits into

sgl-project:mainfrom

MarloweAI:perf/glm53-stack-shared-index-reload

Nov 3, 2026

Merged

[Perf] [MI355X] glm53: Shared layer and MoE hand-offs, NextN entry, K-split front, indexer launch, MoE expert dedup and weight loads, kernarg reload, LM-head loads#85
jhinpan merged 25 commits into
sgl-project:mainfrom
MarloweAI:perf/glm53-stack-shared-index-reload

Commits on Nov 1, 2026
- glm53: Quantize each mid block to FP8 once, for every down task to read
  Hassan Eslami
  committed
  0255109
- glm53: Above 4 rows load the split tiles' KV cache rows ahead of q_b and uk
  Hassan Eslami
  committed
  e2bdc09
- glm53: At 4 rows run one row's up/gate jobs a CTA, the first after that row's first picks
  Hassan Eslami
  committed
  29a6f23
- glm53: Run the NextN entry in four K parts, past the index projection tiles
  Hassan Eslami
  committed
  4929c09
- glm53: Run steps of up to 8 rows on the kernels at TP4, as at TP8
  Hassan Eslami
  committed
  41272ef
- glm53: Load a NextN launch's pointer arguments from the kernarg segment at each stage
  Hassan Eslami
  committed
  76d2048
- glm53: time_layers.py times the NextN launch kinds
  Hassan Eslami
  committed
  4821b8a
- glm53: Split the front's K between a tile's CTA and a spare one, over every CTA
  Hassan Eslami
  committed
  c76bd60
- glm53: Reduce a top-8 round's expert ID only when its maximum is tied
  Hassan Eslami
  committed
  bec4e05
- glm53: Multiply the router's row unnormalized and scale its logits by the rstd
  Hassan Eslami
  committed
  47acd73
- glm53: Above 4 rows score two rows a wave and stage a long run's scores
  Hassan Eslami
  committed
  dcd4e4e
- glm53: Run the index projection tiles alone, and score past the q_norm CTAs
  Hassan Eslami
  committed
  c364faa
- glm53: Stage split's q and new rows before the selection, and warm the selection's code
  Hassan Eslami
  committed
  bb4bf76
- glm53: Above 4 rows stage the scoring's queries in MFMA operand order
  Hassan Eslami
  committed
  3d27d5c
- glm53: The selection's code warm reads no further than the kernel's code end
  Hassan Eslami
  committed
  5ab4f3b
- glm53: Above 4 rows run each expert the step uses once in up/gate and down
  Hassan Eslami
  committed
  aad003c
- glm53: Above 4 rows down skips the MFMAs of a weight batch past the step's expert groups
  Hassan Eslami
  committed
  0a73541
- glm53: Shuffled FP8 weights load 16 bytes a lane, halves swapped between lane rows where read
  Hassan Eslami
  committed
  14f99f8
- glm53: check_layer.py and time_layers.py --similar, a request's rows routing alike
  Hassan Eslami
  committed
  2216161
- glm53: At 4 rows an expert several rows picked streams its up/gate tile once
  Hassan Eslami
  committed
  8a4e87b
- glm53: Load the weights a launch reads once nontemporal
  Hassan Eslami
  committed
  0baa198
- glm53: Load an indexer or dense-MLP launch's pointer arguments from the kernarg segment at each stage
  Hassan Eslami
  committed
  8ae96b6
- glm53: check_layer.py --dump digests the tail's gathered logits
  Hassan Eslami
  committed
  4399689
- glm53: Keep the tail's weight loads in flight across its tiles, the rows' loads ahead of them
  Hassan Eslami
  committed
  b05282e
- glm53: Define the up/gate tile helpers once and fold reload_args' guard into it
  Hassan Eslami
  committed
  b7ae7cb

Commits

Commits on Nov 1, 2026