ROCm: validate native gfx1151 serving on Strix Halo - #260

Open

dbourdea wants to merge 264 commits into

FlashML-org:mainfrom

dbourdea:amd-rocm-gfx1151

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

Commits on Aug 29, 2026

docs(rocm): add warm LAN-223 TPS and GGUF cache reuse
David
committed
6dfa707
docs(rocm): compare Gemma HIP throughput with llama.cpp Vulkan
David
committed
03bec42
docs(rocm): compare FreeToken with llama.cpp ROCm 10
David
committed
4458a85
perf(rocm): target GGUF HIP extensions to active gfx
David
committed
d3a2f57
perf(rocm): retain conservative GGUF math flags
David
committed
a7e77b7
docs(rocm): record LAN-223 TPS optimization results
David
committed
c867f62
docs(rocm): record profiler limitations
David
committed
2c12e6c
tools(rocm): capture reproducible LAN-223 baselines
David
committed
bfd0dde
fix(tools): resolve sibling FreeToken virtual environment
David
committed
ff2a8cc
fix(rocm): profile PyTorch with its wheel SDK
David
committed
5c25f1b
perf(rocm): process two MoE rows per quantized block
David
committed
a6c41e7
Revert "perf(rocm): process two MoE rows per quantized block"
David
committed
0fe9e28
docs(rocm): record rejected MoE geometry experiment
David
committed
652b53a
perf(rocm): test MoE route kernel residency
David
committed
b96257e
Revert "perf(rocm): test MoE route kernel residency"
David
committed
2a25953
docs(rocm): record rejected occupancy experiment
David
committed
43829ee
docs(rocm): record repaired LAN-223 baseline
David
committed
422a1d6
perf(rocm): align Q4 vector packed loads
David
committed
b8f0d8a
Revert "perf(rocm): align Q4 vector packed loads"
David
committed
20d8075
docs(rocm): record rejected Q4 load experiment
David
committed
5f68e1e
perf(rocm): test Q4 MoE FP32 intermediates
David
committed
0cc5e77
docs(rocm): record rejected FP32 MoE experiment
David
committed
ec230c4
Revert "perf(rocm): test Q4 MoE FP32 intermediates"
David
committed
1be4c83
bench(rocm): add Gemma Q4 MoE kernel probe
David
committed
45dfb43
perf(rocm): specialize Q4 MoE two-row wave
David
committed
b2e4271
docs(rocm): record accepted two-row MoE wave
David
committed
bac5ed1
bench(rocm): add Gemma dense Q4 kernel probe
David
committed
a0a4932
perf(rocm): specialize dense Q4 two-row wave
David
committed
56cfadc
docs(rocm): record rejected dense Q4 wave
David
committed
c03fd57
Revert "perf(rocm): specialize dense Q4 two-row wave"
David
committed
5e9ed27
bench(rocm): probe dense Q4 FP32 output
David
committed
f9b41e0
Revert "bench(rocm): probe dense Q4 FP32 output"
David
committed
9f95e6c
docs(rocm): record rejected dense FP32 probe
David
committed
9dda5ef
perf(rocm): constrain dense Q4 GEMV occupancy
David
committed
b5975d8
Revert "perf(rocm): constrain dense Q4 GEMV occupancy"
David
committed
5461b28
docs(rocm): record rejected dense launch bounds
David
committed
f1cd13a
perf(rocm): test indexed Q4 dense HIP kernel
David
committed
05b13b6
Revert "perf(rocm): test indexed Q4 dense HIP kernel"
David
committed
70ac13b
docs(rocm): record rejected indexed Q4 candidate
David
committed
2beadb0
perf(rocm): group Q4 MoE routes per workgroup
David
committed
50a8ae6
Revert "perf(rocm): group Q4 MoE routes per workgroup"
David
committed
d934215
docs(rocm): record rejected MoE route grouping
David
committed
4cd975a
bench(rocm): probe Gemma GQA decode tile
David
committed
d3a8c9b
bench(rocm): parameterize GQA attention probes
David
committed
43d350e
bench(rocm): validate Gemma attention warp candidate
David
committed
2b6a139
docs(rocm): reject GQA warp tuning candidate
David
committed
c07c45f
perf(rocm): scalarize dense Q4 dot temporaries
David
committed
c49938e
Revert "perf(rocm): scalarize dense Q4 dot temporaries"
David
committed
20e6c1a
docs(rocm): record scalarized Q4 candidate result
David
committed
d8f5e99
perf(rocm): reuse Q8 words in two-row MoE
David
committed
18445f0
Revert "perf(rocm): reuse Q8 words in two-row MoE"
David
committed
2a9c264
docs(rocm): reject Q8 reuse MoE candidate
David
committed
42d5b79
bench: expose explicit KV token capacity
David
committed
e13eac5
fix(bench): use server num-tokens flag
David
committed
6cb6f39
docs(rocm): record full cache validation
David
committed
1a0033f
docs(rocm): add current llama control
David
committed
0745a53
docs(rocm): reconcile current operating guidance
David
committed
b8e48bd
docs(rocm): record current-main API revalidation
David
committed
22bcf7f
perf(rocm): use vendored Triton MoE router
David
committed
10029ed
fix(rocm): retain exact MoE router fallback
David
committed
d474c49
docs(rocm): record exact Qwen router revalidation
David
committed
319cc6f
docs(rocm): record current trace and dense Q4 rejection
David
committed
52ea287
docs(rocm): record Q6 scheduling candidate result
David
committed
84c0eb8
docs(rocm): record LAN-223 I/O interference qualifier
David
committed
36864df
tools(rocm): capture LAN-223 I/O interference evidence
David
committed
6c6198b
docs(rocm): record current review validation
David
committed
14986a7
docs(rocm): add clean host llama comparison
David
committed
3875f2f
docs(rocm): record rejected HIP Q4 launch trials
David
committed
061d8e5
docs(rocm): record dense q4 four-wave trial
David
committed
32c7962
docs(rocm): record dense q4 two-wave trial
David
committed
26a376c
perf(hip): test rdna3 sudot4 intrinsic
David
committed
5f040ba
docs(rocm): record accepted rdna3 intrinsic matrix
David
committed
d8abfbd
bench(rocm): add LAN-223 Qwen replication harness
David
committed
96dc476
fix(bench): separate Qwen quality and TPS modes
David
committed
d6ee8ce
fix(bench): harden LAN-223 Qwen recovery
David
committed
91a7418
bench(rocm): add reproducible Qwen scheduler workload
David
committed
2ab83d1
perf(rocm): enable Triton Qwen MoE routing
David
committed
f5d1956
docs(rocm): record Qwen router optimization matrix
David
committed
9599c71
fix(rocm): retain exact Qwen router on HIP
David
committed
c817578
test(rocm): gate Qwen router changes with AIME hash
David
committed
7ad5ca5
bench(rocm): record Qwen quality-gated TPS
David
committed
29eb2bc
perf(rocm): document Qwen FP8 kernel investigation
David
committed
3ae94d8
perf(rocm): screen quality-preserving FP8 GEMV variants
David
committed
1ebdc08
bench(rocm): profile Qwen decode kernel limits
David
committed
0669f01
Merge upstream/main into amd-rocm-gfx1151
David
committed
4897381
docs(rocm): validate current-main Qwen service
David
committed
54d3433
bench(rocm): measure Qwen expert cache copies
David
committed
3856edc
docs(rocm): record live Qwen clock telemetry
David
committed
a8dbb1c
feat(rocm): add reusable gfx1151 kernel cache
David
committed
f6e82b0
docs(rocm): record cache-only scheduler TPS
David
committed
d80a83c
test(rocm): verify gfx1151 cache loads without JIT
David
committed
096fdc1
docs(rocm): record rejected eight-wave gemv test
David
committed
f2f72d9
docs(rocm): record gemv pipeline-depth screen
David
committed
efab1a0
feat(rocm): expose read-only MoE cache diagnostics
David
committed
1bb949e
test(rocm): screen fused expert copy grid
David
committed
6a50d8a
docs(rocm): record gfx1151 FP8 library gate
David
committed
b93cc3e
docs(rocm): record direct HIP GEMV rejection
David
committed
0fb8fa7
docs(rocm): record LAN-223 performance policy audit
David
committed
b722a2b
fix(rocm): detect HIP host extension builds
David
committed
54d6ab2
bench(rocm): preserve harness artifact contract for DPM test
David
committed
5fccc50
test(rocm): protect DPM benchmark artifact handoff
David
committed
68e4a68
docs(rocm): record DPM benchmark evidence contract
David
committed
c1898ab
docs(rocm): record LAN-223 high DPM result
David
committed
163edb4
bench(rocm): add isolated Qwen llama.cpp control
David
committed
9eba13d
docs(rocm): record matched Qwen llama.cpp control
David
committed
f19c385
docs(rocm): scope exact Qwen GGUF port
David
committed
af4e0af
feat(gguf): map Qwen3.5 metadata and Q4_K
David
committed
1e7b6f4
feat(gguf): add mixed Qwen expert banks
David
committed
a567999
feat(gguf): load Qwen dense packed tensors
David
committed
5b3bd5c
feat(gguf): detect Qwen Q6 expert layers
David
committed
7f51bab

Commits on Aug 30, 2026

feat(gguf): run Qwen Q6 expert down layers
David
committed
bce5ef6
build(hip): precompile Qwen mixed GGUF cache rows
David
committed
d1dd473
fix(gguf): preserve direct Qwen RMSNorm scales
David
committed
fcdff54
fix(gguf): load Qwen norm scales without adjustment
David
committed
4c63bc1
fix(gguf): decode Q4_K interleaved nibbles
David
committed
dede23e
fix(gguf): recover Qwen GDN decay logarithm
David
committed
04296a3
fix(gguf): restore Qwen GDN value head order
David
committed
cd95930
fix(gguf): restore Qwen GDN value head projections
David
committed
20379e5
test(qwen): add raw prompt quality control
David
committed
4c61650
fix(completions): honor explicit special token control
David
committed
0ae5be0
fix(gguf): preserve user-defined special tokens
David
committed
c130ed0
test(lan223): add reproducible qwen gguf raw control
David
committed
e37b178
test(lan223): add matched llama raw quality control
David
committed
378f4fd
docs(lan223): record matched qwen q4 raw control
David
committed
9849e2a
feat(rocm): gate native triton moe router for validation
David
committed
5a843a9
perf(rocm): enable validated triton moe router by default
David
committed
5999e56
test(lan223): add isolated gemma4 gguf text control
David
committed
4f0042d
test(lan223): capture gemma4 gguf text quality
David
committed
d61772b
docs(lan223): record gemma4 q4 text control
David
committed
0d87be8
feat(gemma4): map sibling projector tensors
David
committed
c2e8d35
fix(gemma4): parse projector layer names
David
committed
b2b17da
feat(gemma4): load sibling projector for vision
David
committed
60a16a8
feat(gemma4): derive GGUF vision config from projector
David
committed
5168e87
fix(gemma4): use documented projector position table capacity
David
committed
83782cc
fix(lan223): require ready Qwen health before controls
David
committed
592e178
fix(lan223): expand optional vision environment safely
David
committed
cc43a09
feat(server): preserve shaped tensors across message wire
David
committed
dcc65c4
feat(gemma4): add safe OpenAI image preprocessing
David
committed
3a706d2
feat(gemma4): carry OpenAI image inputs to ROCm engine
David
committed
fe9d73f
test(lan223): capture Gemma4 image API control
David
committed
97a0607
test(lan223): run image control before Qwen recovery
David
committed
6877828
docs(lan223): record Gemma4 vision API control
David
committed
ff696d3
fix(lan223): wait for Gemma GPU release before recovery
David
committed
f0fba60
test(lan223): extend Gemma4 visual quality controls
David
committed
f590a81
fix(lan223): verify and retry Qwen recovery
David
committed
e647484
docs(lan223): record multi-case Gemma4 image controls
David
committed
093f06d
test(lan223): add matched llama.cpp Gemma4 vision control
David
committed
d2f3d63
fix(lan223): allow llama Gemma thought channel to finish
David
committed
54b71aa
docs(lan223): record matched Gemma4 vision control
David
committed
07792b6
fix(lan223): wait for Qwen recovery status
David
committed
af9e206
docs(lan223): record Qwen recovery contract
David
committed
d63f52e
test(lan223): capture streamed Gemma vision telemetry
David
committed
5d73f05
test(lan223): stream matched Gemma vision controls
David
committed
f3f7d14
test(lan223): add quality-gated visual TPS verifier
David
committed
506cead
test(lan223): run long visual TPS control
David
committed
f1d8f75
test(lan223): run matched llama visual TPS control
David
committed
b200ad0
fix(lan223): allow llama visual answer after reasoning
David
committed
9e5d33b
test(lan223): extend llama visual reasoning cap
David
committed
5a0b24b
docs(lan223): record long visual API boundary
David
committed
14b6b5d
test(lan223): add tail-latency validation foundation
David
committed
2a3d73f
test(lan223): add versioned Qwen quality suite
David
committed
6600ede
test(lan223): capture runtime benchmark provenance
David
committed
a7e287a
docs(lan223): record clean NVFP4 baseline evidence
David
committed
0f7ad15
test(lan223): add bounded multi-turn tail control
David
committed
a3806a1
docs(lan223): record bounded multi-turn control
David
committed
81b639c
test(lan223): add long-context retrieval control
David
committed
eef1ca0
test(lan223): retain long-context SSE diagnostics
David
committed
fe8f517
test(lan223): distinguish long-context cache hits
David
committed
a7d0f91
fix(lan223): reserve validated Qwen context capacity
David
committed
efca3b0
test(lan223): add concurrent tail-latency control
David
committed
b6f4397
fix(lan223): make concurrent control self-contained
David
committed
a515aaf
docs(lan223): record concurrent tail-latency matrix
David
committed
1865cf4
docs(lan223): record Gemma swap guard recovery
David
committed
901b694
test(lan223): add repeated multi-turn battery
David
committed
0beac25
fix(lan223): tolerate one swap bookkeeping page
David
committed
b17b427
test(lan223): recheck swap after battery health gate
David
committed
c2ca74e
docs(lan223): record multi-turn endurance boundary
David
committed
f63a401
test(lan223): add bounded swap battery diagnostic
David
committed
a5e0091
docs(lan223): record bounded multi-turn characterization
David
committed
535e4a4
test(lan223): add guarded time-share llama control
David
committed
dadf68f
feat(rocm): validate Gemma vision and LAN-223 controls
David
committed
42e5be2
docs(rocm): record full-context MoE cache telemetry
David
committed
5c7f0fd
docs(rocm): add exact Q4 GGUF comparison
David
committed
d8a2dd6
docs(rocm): record Q4 long-context qualification
David
committed
e85645f
fix(rocm): harden Q4 server lifecycle
David
committed
5e41923
test(rocm): scope Q4 endurance swap guard
David
committed
78835aa
test(rocm): summarize Q4 endurance artifacts
David
committed
0fd8cdc
test(rocm): cover Q4 endurance evidence gates
David
committed
233587e
docs(rocm): record Q4 wall-clock endurance
David
committed
663aa9e
docs(rocm): verify post-battery service recovery
David
committed
a937862

Commits on Aug 31, 2026

docs: publish anonymized AMD Strix Halo white paper
David
committed
5a30abf
docs(rocm): distinguish stable and high-cache Q4 gaps
David
committed
cc3a531
test(rocm): allow isolated Q4 candidate sources
David
committed
ce68645
fix(rocm): isolate recovery server lifecycle
David
committed
07b99da
fix(repro): remove hostname from redacted manifests
David
committed
adc76be
test(repro): preserve manifest redaction check
David
committed
5e8448f
fix(repro): accept clean Git worktrees
David
committed
e592130
docs(rocm): add Q4 hardening execution plan
David
committed
70b0ebb
test(rocm): automate Q4 endurance recovery handoff
David
committed
6b49ba0
fix(rocm): accept non-executable recovery scripts
David
committed
eda8b05
fix(rocm): recover clean Q4 endurance candidates
David
committed
9be0790

Commits on Sep 1, 2026

docs(rocm): define Strix Halo performance campaign
David
committed
d54d9f7
docs(rocm): record six-hour trace transition
David
committed
92bf227
perf(rocm): fill RDNA waves in GGUF matvec
David
committed
218104c
Revert "perf(rocm): fill RDNA waves in GGUF matvec"
David
committed
0a1b709
docs(rocm): record rejected Q4 wave candidate
David
committed
fb96d93
perf(rocm): add opt-in HIP math candidate
David
committed
dd8bc3b
Revert "perf(rocm): add opt-in HIP math candidate"
David
committed
44e8b78
docs(rocm): record rejected HIP math candidate
David
committed
0695a77
perf(rocm): screen wider K-quant vector work
David
committed
921ec3f
Revert "perf(rocm): screen wider K-quant vector work"
David
committed
0c17396
docs(rocm): record rejected K-quant vector candidate
David
committed
c93ead6
perf(rocm): screen K-quant two-row MoE vectors
David
committed
3ffb1c6
Revert "perf(rocm): screen K-quant two-row MoE vectors"
David
committed
3873d55
docs(rocm): record rejected two-row K-quant candidate
David
committed
e62f74b
perf(rocm): screen wider Q8 vector dot
David
committed
dea5d6f
Revert "perf(rocm): screen wider Q8 vector dot"
David
committed
443bb25
docs(rocm): record rejected Q8 vector candidate
David
committed
d31f0a9
docs(rocm): define modern MMVQ replacement lane
David
committed
9db166e
bench(rocm): add exact Qwen K-quant MoE kernel screen
David
committed
b38dfc9
docs(rocm): record exact Qwen MoE microbaseline
David
committed
cb5af6b
docs: use GMKtec EVO-X2 public naming
David
committed
95c63c3
docs(rocm): gate modern MMVQ candidate by source audit
David
committed
ad9f870
docs(rocm): record rejected decode-wave screen
David
committed
f5f4419
docs(rocm): verify Qwen recovery after C06
David
committed
8472354
docs(amd): use GMKtec EVO-X2 benchmark naming
David
committed
b3ea5dc
docs(rocm): reject unchanged upstream router policy
David
committed
ac9f6f4
docs(rocm): qualify current Triton router path
David
committed
ffac13c
docs: record router API qualification matrix
David
committed
66e1227
docs: record upstream cache safety gate
David
committed
ac04029
docs: record baseline fused MoE ROCm fault
David
committed
a97ce69
fix(rocm): correct grouped MoE second projection stride
David
committed
04a742f
fix(rocm): use staged grouped MoE alignment
David
committed
9770845
test(rocm): cover grouped MoE alignment ownership
David
committed
a874251
docs(rocm): record fused moe repair and cache gate
David
committed
323dc0f
fix(scripts): mark AMD runner helpers executable
David
committed
9ddf0b4
fix(bench): decouple scheduler harness source
David
committed
ce86d92
feat(bench): parameterize isolated Q4 graph capture
David
committed
c66d4d2
docs(bench): record cache and graph campaign evidence
David
committed
2e7c323
docs(bench): record graph replay qualification
David
committed
4f521ff
feat(rocm): parameterize reviewed GGUF MMV row grouping
David
committed
031e7f2
fix(bench): isolate Q4 candidate extension caches
David
committed
60c3653
docs(bench): record MMV row-grouping screen
David
committed
73131b2
bench(rocm): report client-visible prefill throughput
David
committed
f1baf13
docs(bench): record C29 client prefill result
David
committed
9a62d92
fix(scripts): make ROCm profiler wrapper executable
David
committed
6fd02d1
docs(bench): record ROCprof lifecycle findings
David
committed
f9bb080
feat(bench): add safe ROCprof Q4 trace controller
David
committed
8bf01e9
fix(bench): wait for Q4 scheduler readiness
David
committed
f6d74a1
docs(bench): record finalized Q4 ROCprof trace
David
committed
9ac0c97