Repository navigation
Conversation
… from IRON#169 (CI run 42427147350)
Vendors libgemma4e_dflash and libgemma4e_dflash_flash (xrt + hrx, .so/.dll/.lib) and src/xclbins/Nemotron-3.5-Bolt-E4B-{,DFlash-}NPU2/dequant_mm_bfp.xclbin,
all built by the IRON PR check run for FastFlowLM_IRON#169 (head eec8caeaf). nemotron3.5-bolt:e4b gets the block-float prefill kernel (CI prefill +42% at 1k,
+38% at 4k, +29% at 16k); nemotron3.5-bolt-dflash:e4b gets the whole gemma4-it:e4b prefill/vision/audio/decode work; nemotron3.5-bolt-dflash-flash:e4b gets the
vision/audio/decode host fixes (its flash prefill is unchanged). Accuracy benches unchanged. Docs: prefill-speed notes on the Bolt and Bolt-DFlash cards.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
…e the block-float prefill speedups Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Vendors the engine libs and the block-float prefill xclbin built by the IRON PR check for AMD-ROCm-Internal/FastFlowLM_IRON#169 (head eec8caeaf, CI run 42427147350, all checks green) so the three Nemotron-3.5-Bolt-E4B models get the speedups
gemma4-it:e4bgot in #974.libgemma4e_dflash(nemotron3.5-bolt-dflash:e4b): the full prefill / vision / audio / decode work; with a draft loaded the 18 KV-shared layers run up to layer 35 over the prompt (the draft reads that layer) and the last six as one decode step.libgemma4e_dflash_flash(nemotron3.5-bolt-dflash-flash:e4b): vision / audio / decode host fixes only; its flash prefill is unchanged (it is the faster choice for short prompts).src/xclbins/Nemotron-3.5-Bolt-E4B-NPU2/dequant_mm_bfp.xclbinand.../Nemotron-3.5-Bolt-E4B-DFlash-NPU2/dequant_mm_bfp.xclbin(same kernel asgemma4-it:e4b).nemotron3.5-bolt:e4balready ran the newlibgemma4e_npu(Gemma 4 E4B: faster prefill, decode, image and audio input (vendor IRON#167 engine libs and block-float xclbin) #974).Results
CI runner,
nemotron3.5-bolt:e4bprefill +42% (1k), +38% (4k), +29% (16k), decode unchanged. Local Strix Halo box, TTFT before -> after: bolt 1k 1.45 -> 0.83 s, 4k 3.57 -> 2.51 s; bolt-dflash 64 tokens 1.18 -> 0.84 s, 1k 2.39 -> 1.38 s, 4k 6.21 -> 4.14 s; image start-up -20% (bolt, bolt-dflash), audio -15..40%.Accuracy unchanged (GSM8K, 30-item battery, needle, WER vs Whisper).
Tested on exactly these vendored files
Accuracy A/B, image/audio/WER, 20+ min stress (chunk-boundary sweep up to 8k, 4 mixed clients, disconnects, over-limit), 300-request soak, multi-slot 12 agents x 4 turns, cold first request: 0 errors, 0 mismatches on all three models (flash: only clean 400s above its ~1k prompt cap).
🤖 Generated with Claude Code