feat(v41): stage text loader and prefill composite integration - #258

Open

hashiqiqixian wants to merge 78 commits into

hw-native-sys:mainfrom

hashiqiqixian:feat/v41-model-entry

Commits on Sep 22, 2026

feat(v41): add model metadata and text tokenizer entry
ChenShenAi
committed
b541f19
feat(v41): add selective checkpoint loading and native weight packing
ChenShenAi
committed
65b7472
feat(v41): prepare layer ownership and logical rank execution plans
ChenShenAi
committed
a0df435

Commits on Sep 23, 2026

feat(v41): define composite executor and runner lifecycle contracts
ChenShenAi
committed
39ddff0
feat(v41): prepare bounded embeddings from TP vocabulary shards
ChenShenAi
committed
453461b
feat(v41): track chunked prefill and request-owned cache state
ChenShenAi
committed
399c729
feat(v41): continue decode from committed prefill state
ChenShenAi
committed
deab8db
feat(v41): orchestrate rank-owned weights and complete layer composites
ChenShenAi
committed
820ff9c
feat(v41): return validated logits to the shared generation loop
ChenShenAi
committed
927e6aa
feat(v41): connect model loading and shared serving lifecycle
ChenShenAi
committed
e09ce20
test(v41): supply explicit greedy sampling parameters
ChenShenAi
committed
62c00b2

Commits on Sep 27, 2026

Add bounded V4.1 SWA and packed-FP4 MoE device segment
ChenShenAi
committed
2a2e127
Fix V4.1 cross-layer communication window lifecycle
ChenShenAi
committed
323bac6

Commits on Sep 28, 2026

Add real-weight SWA/MoE bundles and two-layer numerical validation
ChenShenAi
committed
a3b2d5f
Use lib MoE heap budget in the real-weight diagnostic
ChenShenAi
committed
1633399
docs: register the V4.1 segment guide in navigation
ChenShenAi
committed
b52934f
Fix lib reference comparator invocation and allow saved-output checks
ChenShenAi
committed
2bfcdcf
Add post-run half-layer reference diagnostics without device feedback
ChenShenAi
committed
7ef97ca
docs: record real-weight memory and numerical validation boundaries
ChenShenAi
committed
0f59700
Keep stage and end-to-end numerical gates distinct and mandatory
ChenShenAi
committed
c257c93
Add CPU trace replay to locate the first accumulated precision divergence
ChenShenAi
committed
d0a7e17
Trace attention projection and quantization divergence on saved inputs
ChenShenAi
committed
3b06edc
Add saved-boundary interventions for precision bisection
ChenShenAi
committed
e33cc07
Allow selective internal tensor capture for precision diagnosis
ChenShenAi
committed
9d9ce9a
Support first-layer reference capture for matching device dumps
ChenShenAi
committed
5db8442
Preserve per-layer runtime dumps for dataflow bisection
ChenShenAi
committed
bc0023a
Fix dump traversal before moving completed layer artifacts
ChenShenAi
committed
466b5c0
Add independent attention reference precision sensitivity replay
ChenShenAi
committed
83f7a17
Validate SWA chain with checkpoint embedding input and replayable initial state
ChenShenAi
committed
80eb504
test(v41): allow explicit token IDs in embedding precision control
ChenShenAi
committed
8d02950
fix(v41): compare decoded FP8 values in precision diagnostics
ChenShenAi
committed
6b4b873
docs(v41): record text embedding precision control and remaining failure
ChenShenAi
committed
70191be
test(v41): adopt DSV4 layer residual budget and retain legacy profile
ChenShenAi
committed
85f3362
test(v41): record approved accumulated pre-mix budget
ChenShenAi
committed
f74847b
test(v41): probe FP8 projection reference accumulation sensitivity
ChenShenAi
committed
f46d1f8
test(v41): isolate attention RMSNorm reference reduction order
ChenShenAi
committed
7aa7918
feat(v41): pack fresh HC inputs into request-aware TP token slabs
ChenShenAi
committed
71e9929
feat(v41): lower private SWA pages into causal chunk metadata
ChenShenAi
committed
b5f07cd
feat(v41): wire checkpoint RoPE and causal request pages into SWA validation
ChenShenAi
committed
2ec6fd8
fix(v41): replicate indexer heads required by current lib composites
ChenShenAi
committed
98a4311
feat(v41): dispatch prefill composites with independent communication epochs
ChenShenAi
committed
78c0c52
feat(v41): prepare producer-owned compressed attention weight bundles
ChenShenAi
committed
32641e8
feat(v41): lower compressed pages and stable compressor state metadata
ChenShenAi
committed
b032b29
feat(v41): bind resident prefill producers across layer modes
ChenShenAi
committed
589c548
test(v41): validate real-weight C2A producer and reuse chain
ChenShenAi
committed
6f52516
test(v41): trace precision boundaries in each DP partition
ChenShenAi
committed
fd0b762
refactor(v41): load attention probes without expert payloads
ChenShenAi
committed
9f9a5c3
Fix: retain V4.1 cache ownership across paused requests
ChenShenAi
committed
141c01a
Docs: record real-text precision candidate and page ownership limits
ChenShenAi
committed
85ee35e
Add: validate ragged C2A prefixes and empty DP groups
ChenShenAi
committed
fabf011
Add: validate resident C2A compressor continuation across chunks
ChenShenAi
committed
0355d40
docs: record unresolved projection precision experiment
ChenShenAi
committed
f28588d
docs: record successful real-weight C2A continuation
ChenShenAi
committed
861c36d

Commits on Sep 29, 2026

docs: record isolated V4.1 TP communication mismatch
ChenShenAi
committed
4e83854
docs(v41): record verified communication alignment fix
ChenShenAi
committed
043b7bf
test(v41): prepare real-weight C1A producer chains
ChenShenAi
committed
f3608ea
fix(v41): keep C1A index argument write ranges disjoint
ChenShenAi
committed
a11532c
test(v41): retain per-layer compressed attention dumps
ChenShenAi
committed
14e6a5a
docs(v41): record C1A device coverage and numerical failures
ChenShenAi
committed
8da7adc
docs(v41): record C1A probability narrowing isolation
ChenShenAi
committed
c5945eb
docs(v41): correct C1A diagnostic reference and reject PV candidate
ChenShenAi
committed
77e4844
docs(v41): record Q-A improvement and cache error propagation
ChenShenAi
committed
de14b45
test(v41): isolate C1A reuse attention replay
ChenShenAi
committed
9ad837e
docs(v41): record C1A continuation and six-layer results
ChenShenAi
committed
78530b8
test(v41): preserve native gates for isolated kernel candidates
ChenShenAi
committed
f570f02
docs(v41): record isolated Q-A accumulation findings
ChenShenAi
committed
d19ba71
docs(v41): close Q-A device and FP64 comparison
ChenShenAi
committed
9a60f59
test(v41): cover bounded cross-page prefill continuation
ChenShenAi
committed
2c3c6f7
test(v41): retain empty partitions in cross-page diagnostics
ChenShenAi
committed
4b39b38
docs(v41): record original Reuse Q-A control result
ChenShenAi
committed
52cb5f9
docs(v41): record C2A cross-page device validation
ChenShenAi
committed
fac8eb1
docs(v41): distinguish C1A crossing from validated C2A
ChenShenAi
committed
0f3fb63
docs(v41): record C1A cross-page device validation
ChenShenAi
committed
c562231
Add V4.1 decode backbone and final output boundaries
ChenShenAi
committed
5e7d564
Map selective V4.1 weights into decode source slots
ChenShenAi
committed
a28ce2c
Validate V4.1 decode weight geometry on checkpoint
ChenShenAi
committed
f59b6f3
Check every selected V4.1 decode weight ABI field
ChenShenAi
committed
959f959

Commits on Sep 30, 2026

Validate V4.1 cache families before runner allocation
ChenShenAi
committed
4c7b628