Repository navigation

Repair B200 benchmark correctness, runtime lifecycle, and validation coverage - #21

Merged

cfregly merged 66 commits into

mainfrom

codex/b200-results-hardening-20260906

Sep 7, 2026

Commits on Sep 6, 2026

Reuse pooled cache views and write RoPE directly into cache
cfregly
committed
0058fc6
Own compiled inference outputs and preserve memory regressions
cfregly
committed
9f8548c
Describe single-rank hybrid EP without distributed collectives
cfregly
committed
66e2849
Repair full-sweep cache views, compilation, capture and output lifecycle
cfregly
committed
9aef308
Keep hybrid EP training steps fixed and correct Llama normalization
cfregly
committed
b601e98
Verify every MoE logit and retain B200 numerical diagnostics
cfregly
committed
e227138
Preserve Llama eager rounding within max-autotune compilation
cfregly
committed
cc58ab1
Allocate grouped MoE host metadata before compilation
cfregly
committed
555713f

Commits on Sep 7, 2026

Validate full transfer destinations and meaningful inference outputs
cfregly
committed
aafa2c7
Repair worker TLS lifetime and persistent compiled inference state
cfregly
committed
e46514f
Verify complete conditional graph outputs and correct workload documentation
cfregly
committed
fa5a1e8
Keep verification benchmarks live and bind checks to consumed inputs
cfregly
committed
2fcf92c
Align seed guidance with fresh-input verification
cfregly
committed
a48ab02
Respect verification seeds across inference comparison pairs
cfregly
committed
84ba3a0
Capture complete NanoChat requests for CUDA graph replay
cfregly
committed
6f1f686
Bind quantization inputs and release owned compiled graph state
cfregly
committed
1b8caad
Execute two-rank communicator benchmarks and capture live verifier metadata
cfregly
committed
38847c6
Isolate recreated NCCL groups and restore ZeRO-1 weight updates
cfregly
committed
4d29473
Match DDP training math and validate full per-rank child outputs
cfregly
committed
1934034
Preserve harness seeds in benchmark setup prologues
cfregly
committed
a570bc9
Compare Llama compilation using shared stable SDPA arithmetic
cfregly
committed
6c14af3
Verify router outputs against live bounded token inputs
cfregly
committed
670bee6
Verify full quantization outputs and report measured precision data honestly
cfregly
committed
0508653
Preserve active harness seeds through initialized benchmark setups
cfregly
committed
f526cb5
Expose memory-bound gate metadata and verify complete live outputs
cfregly
committed
dad746b
Record full B200 validation and remaining execution gaps
cfregly
committed
6a2a00c
Clear all replicated ZeRO-1 gradients and synchronize initial weights
cfregly
committed
d9acf7d
Deliver target CLI arguments before isolated worker imports
cfregly
committed
8c164cb
Match BF16 optimizer arithmetic across the DDP comparison
cfregly
committed
892fab1
Document two-B200 DDP attribution and sharded optimizer fixes
cfregly
committed
73bba76
Record Llama source speedup and exact two-GPU DDP verification
cfregly
committed
1fa9584
Exercise the production ZeRO-2 communication hook in CPU coverage
cfregly
committed
db05335
Flush partial DDP accumulation groups with correct normalization
cfregly
committed
34d85b7
Preserve native worker stdout before parsing benchmark results
cfregly
committed
40082f1
Transport caller seeds through chapter workers and imported adapters
cfregly
committed
6fb4ec6
Keep CUDA graph tree ownership for explicit compiled graphs
cfregly
committed
260728f
Parse throughput units exactly and retain per-rank summaries
cfregly
committed
54038a8
Transport harness seeds through cache-aware distributed workers
cfregly
committed
954ce67
Use the supported DDP reducer for gradient accumulation
cfregly
committed
b9e0d9e
Preserve caller seeds in topology and GPT model setup
cfregly
committed
93e7f36
Preserve caller RNG streams in direct benchmarks and templates
cfregly
committed
085ed03
Record compiler-policy and distributed validation findings
cfregly
committed
c35f7b0
Align seed hygiene assertions with caller-owned setup
cfregly
committed
a7aa603
Ignore local Hypothesis cache at the repository root
cfregly
committed
9d20a0a
Seed parent references before resolving torchrun launch specs
cfregly
committed
b8a6d2a
Preserve caller seeds in chapter eight benchmark bases
cfregly
committed
8a8eb82
Model fused and eager memory traffic separately
cfregly
committed
e2dca86
Reject missing arithmetic repeats in memory-bound verification
cfregly
committed
6f4e4f4
Preserve fresh setup seeds in shared benchmark implementations
cfregly
committed
766a9e7
Bind cache-aware verification to the live prompt bank
cfregly
committed
42e324e
Provide a reproducible vLLM FlashAttention 4 wheel backport
cfregly
committed
63885e8
Preserve caller seeds across remaining shared benchmark setups
cfregly
committed
48b647d
Expose the validated vLLM attention backend for exact router comparisons
cfregly
committed
3380c56
Exercise the decode post-timing verification lifecycle in the seed control
cfregly
committed
a5e30f8
Record B200 correctness closure and reproducible router validation
cfregly
committed
a36c98c
Keep router prompt allocation outside the measured admission path
cfregly
committed
0fc2322
Honor explicit zero request groups without restoring default work
cfregly
committed
5a47982
Record normal dual-pool results and latest GPU validation
cfregly
committed
5ecf111
Keep the standalone MoE demo default seed outside benchmark setup
cfregly
committed
feec925
Align parity and hygiene tests with caller seeds and complete outputs
cfregly
committed
aae630d
Seed restart and speculative decode parity tests explicitly
cfregly
committed
8c22453
Repair integrated suite fixtures and preserve generated validation guidance
cfregly
committed
9d342cc
Preserve routing validation guidance across all README generation
cfregly
committed
fac47b6
Record complete B200 diagnostic results and passing repair validation
cfregly
committed
f28d63c
Collect only validated metrics for NCU application-range replay
cfregly
committed
d6aee0c
Run distributed optimizer regression workers from the code root
cfregly
committed
d77181e