inference engine for rdna3. codex slop. - #3

Draft

rishavsanjay wants to merge 168 commits into

Commits on Aug 6, 2026

radeon forge: add package skeleton
rishavsanjay
committed
3512801
radeon forge: define workload and trial contracts
rishavsanjay
committed
f5c5c91
radeon forge: add append-only experiment ledger
rishavsanjay
committed
ef8e61e
radeon forge: add hard-gated successive halving tuner
rishavsanjay
committed
af475b6
radeon forge: parse AMD kernel resource metadata
rishavsanjay
committed
9185982
radeon forge: add discoverable kernel-family search spaces
rishavsanjay
committed
e67955c
radeon forge: compile candidate families for gfx1100
rishavsanjay
committed
f568d1d
radeon forge: add executable benchmark protocol
rishavsanjay
committed
8ac0ebe
radeon forge: test contracts metadata and hard-gated tuning
rishavsanjay
committed
0439d06
radeon forge: document architecture metrics and current boundary
rishavsanjay
committed
6b41c83
radeon forge: record RDNA3 kernel corpus and provenance
rishavsanjay
committed
ea87596
radeon forge: add scoped deny-by-default permissions
rishavsanjay
committed
125ae83
radeon forge: test command protocol and permission gates
rishavsanjay
committed
9082329
radeon forge: add native gfx1100 assembly GEMM family
rishavsanjay
committed
cfe1770
radeon forge: add workload package
rishavsanjay
committed
5d5d4e8
radeon forge: add gfx1100 assembly GEMM benchmark oracle
rishavsanjay
committed
fc44f9b
radeon forge: add permissioned gfx1100 tuning CLI
rishavsanjay
committed
5f75d66
radeon forge: add deterministic local knowledge retrieval
rishavsanjay
committed
2e475df
radeon forge: add loopback-only structured local planner
rishavsanjay
committed
1d54bef
radeon forge: test local planner privacy and RAG citations
rishavsanjay
committed
db455b1
radeon forge: add multi-turn proposal and approval state machine
rishavsanjay
committed
f9f86d1
radeon forge: test multi-turn approval state machine
rishavsanjay
committed
38eb2b5
radeon forge: document agent layer and first hardware gate
rishavsanjay
committed
9bd33e2
radeon forge: add unified trace recorder
rishavsanjay
committed
e9dece1
radeon forge: add persistent local backend protocol
rishavsanjay
committed
fb2d888
radeon forge: add permissioned local tools
rishavsanjay
committed
6a40c7a
radeon forge: add agent-native session loop
rishavsanjay
committed
6bb2d25
radeon forge: add unified inference engine
rishavsanjay
committed
0923d1c
radeon forge: export runtime API
rishavsanjay
committed
d1343d2
radeon forge: add agentic profile analysis
rishavsanjay
committed
4f5ba39
radeon forge: export profiling API
rishavsanjay
committed
e11c861
radeon forge: add deterministic development backend
rishavsanjay
committed
1fb12e1
radeon forge: add native tinygrad Llama backend
rishavsanjay
committed
5e2d9fe
radeon forge: add tinygrad MockGPU correctness oracle
rishavsanjay
committed
4c4ecde
radeon forge: add local web application server
rishavsanjay
committed
883251d
radeon forge: add integrated UI shell
rishavsanjay
committed
d13e976
radeon forge: style the local inference UI
rishavsanjay
committed
4463dbd
radeon forge: wire the interactive UI
rishavsanjay
committed
12dc2f6
radeon forge: add content-addressed megakernel workspace
rishavsanjay
committed
d34bd5a
radeon forge: add agent-facing optimization tools
rishavsanjay
committed
3a8d4d7
radeon forge: seed typed megakernel contracts
rishavsanjay
committed
6508d77
radeon forge: export synthesis API
rishavsanjay
committed
7ced140
radeon forge: export oracle API
rishavsanjay
committed
3dc5eea
radeon forge: integrate megakernel workspace and tools
rishavsanjay
committed
41c30f1
radeon forge: expose optimization lab API
rishavsanjay
committed
fd2117c
radeon forge: add megakernel optimization lab UI
rishavsanjay
committed
954ef7b
radeon forge: style megakernel optimization lab
rishavsanjay
committed
61babe4
radeon forge: wire megakernel optimization lab
rishavsanjay
committed
0db6500
radeon forge: serve optimization lab styles
rishavsanjay
committed
71c4c43
radeon forge: load optimization lab styles
rishavsanjay
committed
9129951
radeon forge: record measured kernel durations
rishavsanjay
committed
dc9eec4
radeon forge: attach measured kernel and prefill spans
rishavsanjay
committed
069fbff
radeon forge: add portable single-file optimization recipes
rishavsanjay
committed
7c77b6c
radeon forge: support recipe bundle placeholders
rishavsanjay
committed
4c4449d
radeon forge: export recipe library
rishavsanjay
committed
55f95f0
radeon forge: expose recipe import export tools
rishavsanjay
committed
0fac222
radeon forge: add standalone recipe cli
rishavsanjay
committed
b57e2a8
radeon forge: add portable batch-one recipe example
rishavsanjay
committed
c889ab4
radeon forge: test portable recipe round trips
rishavsanjay
committed
3c5df24
radeon forge: expose portable recipes in engine state
rishavsanjay
committed
bf5b909
radeon forge: add recipe ui endpoints
rishavsanjay
committed
ccb3c44
radeon forge: add portable recipe library ui
rishavsanjay
committed
a2cd1b2
radeon forge: style recipe library ui
rishavsanjay
committed
862ecd5
radeon forge: wire recipe import export ui
rishavsanjay
committed
7b187c6
radeon forge: stream real tinygrad kernel profile events
rishavsanjay
committed
57d4ee4
radeon forge: reason over actual kernel profile ranges
rishavsanjay
committed
1d2a48f
radeon forge: harden backend trace field ingestion
rishavsanjay
committed
6d92836
radeon forge: allow named kernel evidence in traces
rishavsanjay
committed
56277d4
radeon forge: add runtime integration tests
rishavsanjay
committed
9ebc191
radeon forge: add externally measured trace durations
rishavsanjay
committed
c34d637
radeon forge: recover pending tools after denied grants
rishavsanjay
committed
9e0b696
radeon forge: add compatibility-gated hook registry
rishavsanjay
committed
626b050
radeon forge: export runtime hook contracts
rishavsanjay
committed
dd3fcaa
radeon forge: preserve hook descriptors in portable recipes
rishavsanjay
committed
9957b11
radeon forge: define concrete optimization hook targets
rishavsanjay
committed
4f2d4eb
radeon forge: make optimization hooks execution-stage aware
rishavsanjay
committed
f499051
radeon forge: export execution-stage hook contracts
rishavsanjay
committed
76b872f
radeon forge: scope default hooks to execution stages
rishavsanjay
committed
148765a
radeon forge: export stage-aware hook predicates
rishavsanjay
committed
d88be92
radeon forge: retain local runtime fingerprint metadata
rishavsanjay
committed
bb3ee54
radeon forge: expose scripted runtime metadata
rishavsanjay
committed
ee23326
radeon forge: apply stage-specific model hooks with rollback
rishavsanjay
committed
5d945b5
radeon forge: inject active optimization state into model turns
rishavsanjay
committed
bf739f3
radeon forge: add permissioned stage-aware hook deployment tools
rishavsanjay
committed
e8c3e63
radeon forge: connect stage-aware hooks to live sessions
rishavsanjay
committed
83739d7
radeon forge: resolve hooks at live inference stage boundaries
rishavsanjay
committed
e27d645
radeon forge: test stage-specific hook selection and rollback
rishavsanjay
committed
5089b6c
ci: validate Radeon Forge branch
rishavsanjay
committed
7fb9ffd
radeon forge: attribute bottlenecks by execution stage
rishavsanjay
committed
2488ea1
radeon forge: unify direct UI actions with scoped permissions
rishavsanjay
committed
f17ef42
radeon forge: expose permissioned hook activation and rollback UI API
rishavsanjay
committed
c7b1d74
radeon forge: visualize execution-stage optimization hooks
rishavsanjay
committed
82895d1
radeon forge: render and control execution-stage hooks
rishavsanjay
committed
e5c8892
radeon forge: style stage-aware optimization controls
rishavsanjay
committed
3f177c4
radeon forge: fit stage profiler and hook panels
rishavsanjay
committed
fc9655c
radeon forge: test stage-aware kernel profiling
rishavsanjay
committed
f3b0973
radeon forge: test portable execution-stage hook round trip
rishavsanjay
committed
2921d94
radeon forge: add asynchronous local agent jobs
rishavsanjay
committed
3388087
radeon forge: expose live partial generation and incremental events
rishavsanjay
committed
3507232
radeon forge: run local agent turns asynchronously
rishavsanjay
committed
e881ae0
radeon forge: export asynchronous job runtime
rishavsanjay
committed
6e0ac05
radeon forge: stream local agent state through asynchronous UI jobs
rishavsanjay
committed
e4799d4
radeon forge: stream local agent turns in the UI
rishavsanjay
committed
3cc38f0
radeon forge: test live asynchronous agent generation
rishavsanjay
committed
609fb3c
radeon forge: require complete compatibility evidence for deployment
rishavsanjay
committed
5f48f5a
radeon forge: add stage-aware empirical autotuning
rishavsanjay
committed
4b94d4b
radeon forge: export empirical search contracts
rishavsanjay
committed
20d1189
radeon forge: expose empirical W7900 autotuning to agents
rishavsanjay
committed
26ad9ba
radeon forge: carry tuned parameters into active runtime hooks
rishavsanjay
committed
40c914c
radeon forge: apply tuned parameters inside stage hooks
rishavsanjay
committed
1541d95
radeon forge: test hard-gated stage autotuning
rishavsanjay
committed
584a519
ci: cover stage autotuning tests
rishavsanjay
committed
b1a95a9
radeon forge: share stage-specific autotuning spaces
rishavsanjay
committed
78b83c3
radeon forge: test portable stage autotuning knowledge
rishavsanjay
committed
1538592
radeon forge: require transition oracles for stateful hooks
rishavsanjay
committed
a3dcf72
radeon forge: export safe deployment policy
rishavsanjay
committed
744e088
radeon forge: use transition-safe deployment registry
rishavsanjay
committed
1fc7d59
radeon forge: test executable adapters and transition-oracle deployment
rishavsanjay
committed
f482c61
radeon forge: add batched prefill and exact KV reuse accounting
rishavsanjay
committed
79b58f1
radeon forge: add exact KV materialization ledger
rishavsanjay
committed
7b02494
radeon forge: use tested KV reuse ledger in local inference
rishavsanjay
committed
dad3652
radeon forge: test exact KV reuse accounting
rishavsanjay
committed
b1411f8
ci: cover exact KV accounting
rishavsanjay
committed
0fa91c5
radeon forge: advertise native inference-engine capabilities
rishavsanjay
committed
db346c3
radeon forge: add bounded native tool-call stream parser
rishavsanjay
committed
6609ba3
radeon forge: test bounded structured tool parsing
rishavsanjay
committed
0424b87
radeon forge: early-stop on native structured tool calls
rishavsanjay
committed
a02b04d
radeon forge: preserve exact structured tool transcripts
rishavsanjay
committed
2e6be35
radeon forge: test lossless native tool transcripts
rishavsanjay
committed
370d5ca
radeon forge: expose native model-serving stream
rishavsanjay
committed
8c23307
radeon forge: add OpenAI-compatible local serving adapter
rishavsanjay
committed
6d7e6be
radeon forge: hide raw tool protocol from standard streams
rishavsanjay
committed
6cd0466
radeon forge: expose OpenAI-compatible local API
rishavsanjay
committed
3125b97
radeon forge: test OpenAI-compatible local API adapter
rishavsanjay
committed
2af3b95
ci: cover structured tools and OpenAI API
rishavsanjay
committed
fe23afa
radeon forge: render OpenAI tools into local model prompts
rishavsanjay
committed
bd4ff58
radeon forge: inject tool contracts into generic local inference
rishavsanjay
committed
ba7829b
radeon forge: test local tool prompt injection
rishavsanjay
committed
aa7fd3c
radeon forge: add signal-backed cancellation token
rishavsanjay
committed
32aa6bd
radeon forge: preserve resumable state after cancellation
rishavsanjay
committed
c79f8c5
radeon forge: serialize and cancel resident model execution safely
rishavsanjay
committed
a29dc18
radeon forge: add cross-token stop sequence matcher
rishavsanjay
committed
cd28964
radeon forge: test cross-token stop handling
rishavsanjay
committed
9c42981
radeon forge: test stop-buffered text transport
rishavsanjay
committed
0996d82
radeon forge: add tool execution provenance context
rishavsanjay
committed
8b22755
radeon forge: add ROCm counter and ATT capture adapter
rishavsanjay
committed
ccd495f
radeon forge: expose stage-scoped ROCm profiler tools
rishavsanjay
committed
d7d7240
radeon forge: test evidence-preserving ROCm captures
rishavsanjay
committed
08ccc7e
radeon forge: test hardware-grounded profiler findings
rishavsanjay
committed
fb538f5
radeon forge: compare compatible hardware profile captures
rishavsanjay
committed
c1b8a61
radeon forge: test comparable hardware profile deltas
rishavsanjay
committed
0141753
radeon forge: add transformer-block transition oracle
rishavsanjay
committed
d3f3632
radeon forge: test transformer-block oracle harness
rishavsanjay
committed
c6a6665
radeon forge: generate agent-facing candidate scaffolds
rishavsanjay
committed
25ac1ac
radeon forge: test agent candidate scaffold contract
rishavsanjay
committed
0f71f58
radeon forge: add crash-safe local session persistence
rishavsanjay
committed
d53a7af
radeon forge: test local session restart recovery
rishavsanjay
committed
18a6225
radeon forge: add local model worker plugin registry
rishavsanjay
committed
e5fac9d
radeon forge: test local model worker plugins
rishavsanjay
committed
d951dd5
radeon forge: add agent evaluation package
rishavsanjay
committed
52ed99f
radeon forge: add frozen private-agent suite evaluator
rishavsanjay
committed
e0b663d
radeon forge: add local agent-suite benchmark CLI
rishavsanjay
committed
498e5e5
radeon forge: add frozen private coding-agent suite
rishavsanjay
committed
c8b2edb
radeon forge: test frozen agent-loop evaluation
rishavsanjay
committed
142c72e
radeon forge: compare agent suites under quality gates
rishavsanjay
committed
d01342c
radeon forge: test quality-gated agent latency comparison
rishavsanjay
committed
9b35ad5
radeon forge: test agent-metric deployment gate
rishavsanjay
committed
208493b
radeon forge: normalize recognized ATT and SQTT exports
rishavsanjay
committed
5492257