feat: qualify native model swapping on AMD - #3

Closed

dbourdea wants to merge 569 commits into

mainfrom

codex/freetoken-swap-lan215

This pull request is big! We're only showing the most recent 250 commits

Commits on Sep 4, 2026

docs: refresh campaign open-work status
David
committed
f4848dd
docs: consolidate cross-model benchmark matrix
David
committed
f6a20b3
docs: inventory in-scope model routing entries
David
committed
b3748a4
docs: record model payload admission gate
David
committed
87427d7
docs: record qwen35 GGUF parser gap
David
committed
c65dbba
docs: compare qwen35 and qwen35moe schemas
David
committed
1984442
feat: admit dense qwen35 GGUF hybrid models
David
committed
30d7d49
fix: preserve qwen35moe GGUF config width
David
committed
6028ff2
feat: decode qwen35 dense GGUF weights
David
committed
c60f0f5
docs: record qwen35 dense loader milestone
David
committed
c9f0fb0
docs: record qwen35 construction validation
David
committed
b00f3ce

Commits on Sep 5, 2026

docs: record reproducible Q4 MMV_Y=4 promotion candidate
David
committed
e62b3c1
docs: record Gemma4 native text matrix
David
committed
e2d8f6c
docs: record matched Gemma4 llama.cpp control
David
committed
7797d2d
feat: opt in to Gemma prefill overlap
David
committed
f74a3e6
docs: record Gemma prefill overlap rejection
David
committed
111f144
perf: enable Gemma ROCm prefill warmup
David
committed
c70b31d
perf: parameterize Gemma expert memory ratio
David
committed
eaf19ab
docs: reject higher Gemma memory ratio
David
committed
8ac3adf
bench: add Gemma prompt length control
David
committed
6916669
docs: record Gemma prompt scaling
David
committed
713026f
bench: match llama.cpp Gemma prompt scaling
David
committed
14a12e8
docs: record matched long prompt control
David
committed
5341af5
bench: match Gemma concurrent prompt scaling
David
committed
c7cffd5
bench: match llama.cpp concurrent prompt scaling
David
committed
5f8cb0e
docs: record matched Gemma concurrency
David
committed
9291f52
docs: record eight-client Gemma stress
David
committed
41f769e
docs: record two-client Gemma control
David
committed
f68a901
perf: parameterize Gemma scheduler request limit
David
committed
1257759
docs: record Gemma scheduler admission candidate
David
committed
5d86645
docs: record max six Gemma scheduler candidate
David
committed
dbb4e63
docs: consolidate current cross-model benchmark matrix
David
committed
03575af
docs: add machine-readable cross-model manifest
David
committed
2a2d8af
docs: refresh live 284b capacity evidence
David
committed
1b34eef
docs: record same-format Qwen GGUF control
David
committed
3a631a3
docs: record exact-format Qwen raw controls
David
committed
f8a9f91
docs: record warmed Qwen GGUF matrix
David
committed
e6884a2
docs: complete warmed Qwen GGUF comparison
David
committed
a50a68e
docs: update campaign completion audit
David
committed
2087e16
docs: add final AMD campaign report
David
committed
54354ea
docs: refresh upstream review status
David
committed
4eca6fe
docs: add upstream handoff checklist
David
committed
15a018a
docs: record final handoff audit
David
committed
fa8e760
docs: anchor 284b capacity limits to primary paper
David
committed
835ae63
docs: compare paper and local expert bandwidth evidence
David
committed
9c39cdb
docs: gate 284b claims on exact checkpoint identity
David
committed
39ab06b
docs: refresh DeepSeek capacity metadata gate
David
committed
a98c4cf
docs: pin DeepSeek release revision for paper gate
David
committed
3622777
docs: record pinned DeepSeek config identity
David
committed
f79a240
feat: add reproducible DeepSeek metadata capacity gate
David
committed
17a5c77
docs: carry capacity rejection into handoff audit
David
committed
abb64cc
docs: add DeepSeek expert slice metadata
David
committed
6d12d09
feat: add isolated DeepSeek expert slice harness
David
committed
7577fb6
test: validate expert slice metadata selection
David
committed
32f1d84
docs: reconcile DeepSeek capacity gate status
David
committed
d8dc4f5
docs: record real-shape DeepSeek ROCm slice
David
committed
e495f3e
docs: compare real-shape slice bandwidth
David
committed
25b0e12
docs: record larger DeepSeek expert slice
David
committed
78cc702
docs: record cross-shard DeepSeek slice
David
committed
7013365
docs: record expert route churn controls
David
committed
31c44dd
docs: add DeepSeek route transfer projection
David
committed
f91446c
test: align HIP GGUF default flag contract
David
committed
803a3ce
docs: record HIP flag contract validation
David
committed
bd6d4a2
feat(rocm): add four-row GGUF MMV screen
David
committed
ff76ede
docs: close current-branch MMV Y4 qualification
David
committed
33ce651
docs: record current MMV Y4 campaign closure
David
committed
57c57c3
fix: make AIME quality contract explicit
David
committed
50bd6dd
docs: distinguish active AIME quality contract
David
committed
4fb00c9
docs: record AIME contract probe boundary
David
committed
865db9a
docs: record paired MMV Y1 control
David
committed
ce8d1ba
test(rocm): align legacy copy catalog assertion
David
committed
8ad5b63
Merge upstream main into AMD ROCm branch
David
committed
4a9c05f
docs: record upstream sync and ROCm guard validation
David
committed
58eab4e
docs: record native ROCm extension build
David
committed
5a56629
docs: preserve ROCm extension build artifact
David
committed
618ae4f
docs: record FP8 GEMV tile32 repeatability gate
David
committed
5afc1f4
docs: reject FP8 tile32 under Qwen concurrency
David
committed
07bf428
docs: correct warmed tile32 concurrency gate
David
committed
e7a5a92
docs: record exact branch ROCm regression
David
committed
08d344e
feat(rocm): expose opt-in NVFP4 deep-K tile
David
committed
47c5926
docs: record deep-K NVFP4 candidate gate
David
committed
cd897f7
docs: record MMV Y8 infrastructure attempt
David
committed
36b9ec7
docs: record clean-window MMV Y8 attempt
David
committed
7e7f77c

Commits on Sep 10, 2026

feat(daemon): add named model swap profiles
David
committed
94021ec
feat(daemon): wait for profile readiness
David
committed
14f344e
test(daemon): cover profile readiness timeout
David
committed
1f80150
feat(daemon): add shutdown client command
David
committed
c980106
docs: anonymize AMD host names and deployment paths
FreeToken contributor
committed
ccc6295
fix(swap): enforce readiness and document direct routing integration
FreeToken contributor
committed
f88da87
fix(bench): derive operator paths and guard public document privacy
FreeToken contributor
committed
6707cb5
fix(amd): load dense Qwen projections using per-tensor GGUF types
FreeToken contributor
committed
0e1364e
docs: normalize anonymous hardware label and sanitize JSON evidence index
FreeToken contributor
committed
63ecac0
docs: record extended live swap qualification and model checksums
FreeToken contributor
committed
94b034a
test(swap): qualify concurrent routing and idle eviction with private native cache
FreeToken contributor
committed
492a52c
docs: anonymize publication metadata and require explicit benchmark hosts
FreeToken contributor
committed
de23ad6
fix(swap): recover previous launch after replacement spawn failure
FreeToken contributor
committed
899d571
fix(swap): recover failed readiness without overriding newer lifecycle intent
FreeToken contributor
committed
accf7f3
test(swap): add strict streaming disconnect qualification gate
FreeToken contributor
committed
4c35cdb
test(swap): exercise real Linux process rollback and worker cleanup
FreeToken contributor
committed
f9b0453
test(swap): assert termination signals and document Linux recovery evidence
FreeToken contributor
committed
64dcc68
docs(swap): record combined branch validation and remaining completion gates
FreeToken contributor
committed
99bdea1
test(swap): qualify live cancellation and real model failure recovery
FreeToken contributor
committed
64ab64f
docs: record real cancellation and model failure recovery results
FreeToken contributor
committed
c0534c6
docs(swap): establish source-backed native parity matrix
FreeToken contributor
committed
ae566f4
feat(swap): validate native routing policy configuration
FreeToken contributor
committed
6ca47ff
feat(swap): add native inference routing admission
FreeToken contributor
committed
a392af9
feat(swap): evict idle routed engines
FreeToken contributor
committed
24849c6
test(swap): exercise invalid profile group validation
FreeToken contributor
committed
901c0b9
feat(swap): expose router admission metrics
FreeToken contributor
committed
2e4316c
feat(swap): add guarded upstream passthrough
FreeToken contributor
committed
ae7ccdd
feat(swap): reload validated routing catalogs
FreeToken contributor
committed
2865ef0
feat(swap): enforce single-slot persistent residency
FreeToken contributor
committed
0d0cea5
feat(swap): cancel routed inflight requests
FreeToken contributor
committed
1b9c59a
feat(swap): report configured and resident profiles
FreeToken contributor
committed
76c3288
docs(swap): distinguish native routing evidence
FreeToken contributor
committed
667277c
test(swap): exercise real loopback SSE proxying
FreeToken contributor
committed
596cfae
feat(swap): record native proxy timing signals
FreeToken contributor
committed
7124501
feat(swap): configure native upstream timeout
FreeToken contributor
committed
8551430
feat(swap): add safe profile request filters
FreeToken contributor
committed
66e50cd
test(swap): cover native router with real child
FreeToken contributor
committed
d5590f4
fix(swap): protect router management with API keys
FreeToken contributor
committed
a440057
docs(swap): record current daemon regression gate
FreeToken contributor
committed
80bb5dd
docs(swap): add native qualification matrix
FreeToken contributor
committed
76f531d

Commits on Sep 14, 2026

feat(swap): add private router event stream
FreeToken contributor
committed
fffde24
feat(swap): allocate dynamic routed ports
FreeToken contributor
committed
7007dbc
feat(swap): add native management load
FreeToken contributor
committed
56afc86
feat(swap): expose native model list and readiness
FreeToken contributor
committed
e5d0a41
fix(swap): keep concrete paths out of router logs
FreeToken contributor
committed
32d56c7
docs(swap): inventory pinned reference surfaces
FreeToken contributor
committed
10bb1b6
fix(swap): reject unsupported group coexistence
FreeToken contributor
committed
4729992
feat(swap): add private local management UI
FreeToken contributor
committed
0dabbd0
feat(swap): safely watch catalog changes
FreeToken contributor
committed
52aebcf
feat(swap): expose router transfer timing
FreeToken contributor
committed
b6ac3ee
test(swap): cover every routed text endpoint
FreeToken contributor
committed
1ee4499
feat(swap): add native routing benchmark gate
FreeToken contributor
committed
44d8354
fix(swap): restrict maintenance benchmark to Linux
FreeToken contributor
committed
c45a5d0
test(swap): validate native benchmark observations
FreeToken contributor
committed
3cd2082
feat(swap): retain benchmark metric snapshots
FreeToken contributor
committed
b238a7f
test(swap): preserve routed upstream errors
FreeToken contributor
committed
af25b64
test(swap): exercise dynamic ports with real children
FreeToken contributor
committed
a75a6aa
fix(swap): freeze active routing policy on reload
FreeToken contributor
committed
67b0f63
test(swap): cover active group policy reload conflict
FreeToken contributor
committed
984bba9
fix(swap): keep router credentials out of upstream
FreeToken contributor
committed
f572b9c
test(swap): verify auth stripping over loopback
FreeToken contributor
committed
de1885a
fix(swap): preserve engine response headers
FreeToken contributor
committed
cb1710d
test(swap): rotate router keys through catalog reload
FreeToken contributor
committed
1d8af20
docs(swap): separate current and historical evidence
FreeToken contributor
committed
5cd63d3
test(swap): verify benchmark scenario state
FreeToken contributor
committed
2d0390c
feat(swap): capture private benchmark hardware snapshots
FreeToken contributor
committed
4e8b7b0
fix(swap): validate native benchmark catalog
FreeToken contributor
committed
7d9465e
test(swap): require benchmark process observations
FreeToken contributor
committed
079bedd
test(swap): verify benchmark listener cleanup
FreeToken contributor
committed
f2a9c21
docs(swap): retain direct benchmark lifecycle evidence
FreeToken contributor
committed
5ce51c0
docs(swap): record current daemon verification
FreeToken contributor
committed
64af6be
fix(swap): verify resident identity in readiness
FreeToken contributor
committed
9d795ae
test(swap): reject stale args and port in readiness
FreeToken contributor
committed
d9dc7f7
fix(swap): expose stale resident identity
FreeToken contributor
committed
5b31904
fix(swap): report stale children as nonresident
FreeToken contributor
committed
644482b
feat(swap): export resident identity metric
FreeToken contributor
committed
6941425
test(swap): require concrete benchmark identity
FreeToken contributor
committed
b01e77e
feat(swap): record usage-derived benchmark throughput
FreeToken contributor
committed
aa1bf63
feat(swap): expose benchmark decode interval
FreeToken contributor
committed
3e4fbad
fix(swap): exclude cancelled streams from completions
FreeToken contributor
committed
a6d5244
feat(swap): qualify native router cancellation
FreeToken contributor
committed
d01ba57
feat(swap): qualify same-model concurrency
FreeToken contributor
committed
92a3987
feat(swap): qualify native TTL eviction
FreeToken contributor
committed
235d3a2
feat(swap): qualify active catalog reload conflict
FreeToken contributor
committed
5b6eda1
feat(swap): qualify failed-switch rollback
FreeToken contributor
committed
53a96ce
test(swap): require rollback accounting evidence
FreeToken contributor
committed
4dbc56b
fix(swap): bind exact re-adopted residency
FreeToken contributor
committed
c62e3a5
test(swap): cover real child router re-adoption
FreeToken contributor
committed
e9901c0
feat(swap): qualify daemon re-adoption
FreeToken contributor
committed
4961590
docs(swap): refresh draft PR evidence
FreeToken contributor
committed
9d0e9d4
docs(swap): keep PR evidence revision-stable
FreeToken contributor
committed
0033721
feat(swap): qualify persistent capacity protection
FreeToken contributor
committed
37df2bb
test(swap): prove unload one and all semantics
FreeToken contributor
committed
1ac3053
docs(swap): record current daemon suite
FreeToken contributor
committed
b3d5c50
feat(swap): qualify conflicting request drain
FreeToken contributor
committed
bd2b630
docs(swap): align parity classifications with ownership
FreeToken contributor
committed
5944629
fix(swap): filter hop-by-hop response headers
FreeToken contributor
committed
b0fe4ea
fix(swap): stabilize unknown model errors
FreeToken contributor
committed
e1370d3
docs(swap): refresh full suite evidence
FreeToken contributor
committed
ca5cfaf
fix(swap): reserve request IDs before admission
FreeToken contributor
committed
05cb116
fix(swap): cancel abandoned admission waiters
FreeToken contributor
committed
e494236
test(swap): classify legacy generate routing
FreeToken contributor
committed
82be471
fix(swap): cancel requests across owned phases
FreeToken contributor
committed
12c5bb6
docs(swap): align research with native ownership
FreeToken contributor
committed
e990b5c
fix(swap): linearize router readiness probes
FreeToken contributor
committed
cc14820
fix(swap): serialize manual and routed lifecycle
FreeToken contributor
committed
94c12f9
fix(swap): retain manual ownership after disconnect
FreeToken contributor
committed
1af92a7
fix(swap): finish manual transactions after disconnect
FreeToken contributor
committed
96d8e60
fix(swap): quiesce routing before daemon shutdown
FreeToken contributor
committed
c4667b1
fix(swap): coordinate daemon exit with routing
FreeToken contributor
committed
68caf9e
fix(swap): make catalog binding atomic with admission
FreeToken contributor
committed
5b6ab25
test(swap): qualify authenticated control plane
FreeToken contributor
committed
d460906
feat(swap): add alternate model aliases
FreeToken contributor
committed
e19b99d
feat(swap): add browser CORS compatibility
FreeToken contributor
committed
2c51f8d
feat(swap): support pinned API key forms
FreeToken contributor
committed
4626450
feat(swap): enforce concurrency admission limits
FreeToken contributor
committed
c5dee9c
feat(swap): expose exact model load status
FreeToken contributor
committed
744eb0c
fix(swap): align public models alias
FreeToken contributor
committed
2e720ef
feat(swap): support namespaced model ids
FreeToken contributor
committed
0933243
docs(swap): classify streaming load feedback gap
FreeToken contributor
committed
e376fa6
feat(swap): stream cold model loading state
FreeToken contributor
committed
c1b1004
test(swap): gate cold loading with real child
FreeToken contributor
committed
f55dec7
docs(swap): refresh current PR audit
FreeToken contributor
committed
51c1869
feat(swap): preserve stateless response routes
FreeToken contributor
committed
a56ccdc
ci(swap): run daemon suite on Linux
FreeToken contributor
committed
fdfce8b
ci(swap): expose bounded Linux failures
FreeToken contributor
committed
0b24bbf
test(swap): fix Linux SSE fixture
FreeToken contributor
committed
b25b2cb

Commits on Sep 15, 2026

test(swap): emit real SSE newlines on Linux
FreeToken contributor
committed
e906b73
test(swap): tolerate vanished Linux workers
FreeToken contributor
committed
5ee1e26
docs(swap): record hosted Linux process evidence
FreeToken contributor
committed
098d758
ci(swap): report clean Linux totals
FreeToken contributor
committed
22cbd39
fix(swap): gate maintenance by exact host
FreeToken contributor
committed
bf70a04
docs(swap): record maintenance host gate evidence
FreeToken contributor
committed
4615daa
fix(daemon): make AMD memory evidence fail closed
FreeToken contributor
committed
261d80f
docs(swap): record AMD observability evidence
FreeToken contributor
committed
8930738
Add safe model capability metadata
FreeToken contributor
committed
89d026c
Record capability metadata verification
FreeToken contributor
committed
bf4a584
Add safe request parameter filters
FreeToken contributor
committed
70fe9f4
Record request filter verification
FreeToken contributor
committed
2b35abd
feat: add native warm and pin selectors
FreeToken contributor
committed
cbf00d6
docs: record selector verification
FreeToken contributor
committed
6ba35b0
feat: add runtime routing profiles
FreeToken contributor
committed
3b257bd
docs: record runtime profile evidence
FreeToken contributor
committed
efe605e
feat: add safe readiness and proxy targets
FreeToken contributor
committed
c0833f4
docs: record readiness target evidence
FreeToken contributor
committed
743361b
feat(swap): support upstream model name overrides
FreeToken contributor
committed
2c90c60
docs(swap): record upstream model rewrite evidence
FreeToken contributor
committed
8f71070
feat(swap): expose safe model listing metadata
FreeToken contributor
committed
07abf68
docs(swap): record model metadata parity evidence
FreeToken contributor
committed
08d807b
feat(swap): support per-profile upstream timeouts
FreeToken contributor
committed
53b1c6c
docs(swap): record profile timeout evidence
FreeToken contributor
committed
da02637
feat(swap): preload a model through native startup
FreeToken contributor
committed
a3fc0dd
docs(swap): record startup preload evidence
FreeToken contributor
committed
17fe93c
Add pinned source inventory and safe upstream guard
FreeToken contributor
committed
5a98b29
Record source inventory and upstream guard evidence
FreeToken contributor
committed
4b81267
Add bounded activity and opt-in captures
FreeToken contributor
committed
38c2457
Record bounded activity and capture evidence
FreeToken contributor
committed
b2c80e8
feat(swap): persist activity and add session-aware UI
FreeToken contributor
committed
493aecd
docs(swap): record durable activity CI evidence
FreeToken contributor
committed
df13574
feat(swap): add bounded periodic performance history
FreeToken contributor
committed
174bda1
docs(swap): record performance history evidence
FreeToken contributor
committed
da4b8df
docs(swap): record current combined-tree CPU evidence
FreeToken contributor
committed
c5d130b
test(swap): gate live periodic performance evidence
FreeToken contributor
committed
39c3aaa
docs(swap): refresh final CPU qualification evidence
FreeToken contributor
committed
a5846c0
docs: explain FreeToken swap code line by line
FreeToken contributor
committed
0c356c1

Commits on Sep 19, 2026

LAN-215 private combined qualification tree
LAN-215 Qualification
committed
150c88e

Commits on Sep 24, 2026

feat(daemon): qualify native model swapping on AMD
dbourdea
committed
56e0232