- Split Qwen3::forward_decode_paged into decode_prepare (host-side
block allocation + table upload) and decode_core (pure-GPU compute
reading token ids and positions from device buffers via
embedding_device_ids + rope_inplace_device_pos). This makes the
entire Qwen3 decode step CUDA-graph-capturable, mirroring the
gpt_oss.rs architecture.
- Add qwen3_graph.rs: Qwen3DecodeGraph + GraphedQwen3Decoder, a port
of the gpt_oss_graph.rs whole-step capture pattern. Lazy policy:
first decode eager (warms pool + cuBLAS), second captures, rest
replay. Batch>1 always falls back to eager.
- Wire GraphedQwen3Decoder into bench-speculative's draft decode path;
all 4 draft.forward_decode_paged call sites + replay_draft_tokens
now route through the graphed decoder. Per-benchmark caches persist
across prompts for graph reuse.
- Gamma sweep result (10 prompts × 32 tokens, --use-verify-logits):
γ=1 → 0.57×, γ=2 → 0.57×, γ=4 → 0.49×, γ=6 → 0.41×, γ=8 → 0.36×.
All matched=true, verify_decode_mismatches=0.
Acceptance drops sharply with γ (66% → 40% → 25%) because Qwen3-0.6B
is too inaccurate a draft for Qwen3-8B. Speedup still <1.
Current ceiling analysis: verify costs ~13ms (same as one target decode)
so speculative decoding only wins if acceptance × (tokens/round) >>
(draft_cost + verify_cost) / baseline_decode. With this draft model,
the crossover requires either (a) a much smaller verify cost (batch-GEMM
path, which trades correctness), or (b) a fundamentally better drafter
(EAGLE-style heads, or n-gram lookup).
Add launch_gemv_bf16_batched: runs M m=1 GEMVs in a single 3D grid
launch (z = batch row) with numerically identical output to M sequential
launch_gemv_bf16 calls — same K-block partial accumulation, same
fixed-order reduction. Verified on dash5 with 10 prompts × 32 tokens:
matched=true, verify_decode_mismatches=0.
Expose as matmul_batched_gemv(a: [M,K], b: [K,N]) → [M,N] in
xserv-kernels. Replace the old matmul_rows_gemv helper in qwen3
forward_verify_paged_decode_attention; the per-row loop over matmul_2d +
concat_rows is replaced by a single matmul_batched_gemv call that
allocates the partials buffer in one shot and launches 2 kernels instead
of 2*M.
Current speedup_e2e is 0.47× (same ballpark as Phase 23 0.44×);
the batched launch saves ~3 ms overhead but this is small relative to
the total 28 ms spec cost. The path forward (per docs/24 §4) is
higher acceptance rate or cheaper draft, not further kernel optimization.
partial_cmp().unwrap() in the top-k / top-p sort and softmax paths would
panic the engine thread on a single NaN logit. The greedy argmax path
is already NaN-safe. Add a one-pass NaN → -inf sweep on the extracted
last_row before temperature scaling, which is equivalent to masking the
token and keeps the sampler deterministic. Warn once when triggered so
the underlying numeric bug isn't silently hidden.
Interactive REPL used to always call sample_greedy_last on both the
paged and legacy KV paths, so temperature/top-k/top-p and the repetition
penalty added in the sampling module were unreachable from the CLI.
- flag() helper parses --max-tokens / --temperature / --top-k / --top-p
/ --rep-penalty / --rep-window (defaults preserve prior behavior:
temperature 0, top-p 1, penalty 1, window 512).
- pick_next() dispatches to sample_greedy_penalized only when
temperature==0 and rep_penalty>1, otherwise to sample().
- Both Qwen3/GPT-2 paths and the GptOss paged path share the same
sampler and both feed the rolling history window used for the penalty.
- Prompt input now unescapes literal "\n" so multi-turn prompts can be
typed on one line.
Phase 22 lands a correctness-only speculative decoding loop for Qwen3
target + Qwen3 small draft (batch=1, greedy, gamma=4). Phase 23 turns
verify logits into the authoritative acceptance signal so mirror-decode
per accepted token is no longer needed.
- paged_kv_cache: truncate_sequence(slot, new_len) shrinks a registered
sequence, freeing whole physical blocks no longer reachable and
leaving the slot registered. Covered by a CUDA-gated unit test.
- qwen3: forward_verify_paged_decode_attention writes the draft window
into the target cache, runs the same paged decode attention kernel per
draft token, and uses matmul_rows_gemv so linear layers follow the
single-token decode BF16 rounding path.
- bench-speculative: new bench binary drives the state machine with
--gamma / --gen-tokens / --prompts / --use-verify-logits /
--verify-path flash|paged-decode / --dump-verify-mismatches, and
compares baseline vs spec token sequences plus TPOT / tok/s / speedup.
- docs/22 records the decode-authoritative v0 result and dash5 numbers
(matched=true, speedup_e2e ~0.29x, verify_decode_mismatches>0 under
--use-verify-logits).
- docs/23 records the paged-decode verify path (matched=true,
verify_decode_mismatches=0, 50x64 speedup_e2e ~0.44x) and the
next-step performance TODO.
BF16 greedy decode was sensitive to inter-block scheduling when logits
were close, which broke speculative-decoding verify-vs-decode parity.
- gemv.cu: write per-K-block partials, then reduce in fixed block order
in a second kernel instead of atomicAdd across K-blocks. Scratch
buffer size is now n * ceil(k / GEMV_TILE_K); gemv_scratch_elems()
exposes this to callers, and decode_graph.rs sizes fp32_hidden/q/kv/
intermediate/vocab from it.
- paged_attention.cu: replace atomicAdd merge of warp outputs with
per-warp shared partials reduced in warp-id order for both the base
and sinks kernels.
sample() at temperature 0 copied the full [seq, 201088] BF16 logits
to the host and scanned them every token (~1 ms/token). Use the
Phase 15 argmax kernel (block reduction + 4-byte D2H) when logits are
BF16 on GPU; bench-gpt-oss's greedy sampler likewise. Exact-tie
logits may break differently than the host scan — greedy trajectories
can legitimately diverge at a tie token (GSM8K unchanged).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Split forward_decode_paged into host bookkeeping (decode_prepare +
ids/pos upload + advance_seq_len) and a pure-GPU decode_core. The
paged-KV and sparse-MoE designs already read every per-step variable
(block table, context lens, expert ids) from stable-address device
buffers, so decode_core captures as-is.
GptOssDecodeGraph captures lazily on the second decode step (the
first eager step warms cuBLAS) after a "retained warmup": the step
runs once with the allocator quarantine on, stocking the pool with a
dedicated block for every allocation so the capture itself never
pool-misses (a cudaMalloc while capturing is illegal — and the
capture's own quarantine is what would otherwise starve the pool).
NCCL all-reduces capture cleanly; TP=2 replays in lockstep.
Wired into tp_engine, bench-gpt-oss, and xserv-chat via
GraphedGptOssDecoder (batch>1 falls back to eager;
XSERV_DECODE_GRAPH=0 disables). Greedy tokens identical to eager.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Thread-local launch stream (xserv_cuda::stream): every kernel
wrapper, cublasSetStream, and NCCL collective now launches on
current_stream_raw() — the legacy null stream by default (behavior
unchanged), or the capture stream installed via push_stream during
graph capture. Capture is impossible on the legacy stream.
- Allocator retain mode: blocks freed inside a retain window are
quarantined (RetainedBlocks) instead of pooled, so an instantiated
graph keeps exclusive ownership of every intermediate buffer it
references across replays.
- Capture mode GLOBAL -> THREAD_LOCAL: concurrent TP rank threads
must not poison each other's captures with their own cudaMallocs.
- embedding_device_ids / rope_inplace_device_pos: variants reading
token ids / positions from persistent device buffers, replacing the
per-call host upload that a captured region cannot contain.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- --pp with gpt-oss now fails with a clear message instead of a
cryptic missing-weight panic inside the Qwen3-only PP engine.
- Sparse GEMV wrappers assert K%16==0 (FP8) / K%32==0 (MXFP4) — the
uint4-vectorized kernels would silently drop a tail otherwise.
- Document the topk_ids buffer holding i32 under an F32 dtype label
(DType has no I32).
- Drop unused imports/locals and the cuBLASLt scale-mode constants
orphaned by the strided-batched FP8 rework (e631a71).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Decode carried three leftover cudaDeviceSynchronize (prefill one) from
NaN debugging — the Qwen3 path has none and the logits D2H in sample()
already orders against the null stream.
add_bias for rows>1 round-tripped the bias through the CPU (D2H + host
tile loop + H2D) on every call — 96 times per prefill across q/k/v/o.
Replace with a bias_add_2d broadcast kernel.
dash5, FP8 TP=2, warm-server: TTFT 35/49/94 -> 29/42/79 ms
(short/medium/long), TPOT 7.19-7.32 -> 6.99-7.21 ms. Greedy tokens
unchanged; GSM8K-50 94%.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Dense MoE replicated x across all 16 local experts and ran the full
batched GEMM, reading every expert's weights per token; the weighted
sum then discarded 12 of 16 results. Decode is memory-bound, so this
was ~8x wasted expert bytes — the entire decode gap vs llama.cpp.
New fused expert-indexed GEMVs (csrc/moe/moe_sparse.cu) read
topk_ids on-device (no host sync) and early-return block-uniformly
for experts other ranks own. FP8 runs W8A16 (activations stay BF16 —
tensor cores are irrelevant at M=1, and activation quantization error
disappears); MXFP4 runs W4A16. Per-expert bias + scale fused into the
GEMV epilogue; slot-indexed weighted sum skips (never multiplies)
unwritten non-local slots. Dense path retained for num_tokens > 8
(prefill) and via XSERV_DENSE_MOE=1 for A/B.
dash5 (RTX 5090), gpt-oss-20b FP8, TP=2: decode TPOT 13.9 -> 7.6 ms.
Warm-server vs llama.cpp MXFP4 TP=2: TPOT 7.19-7.32 vs 7.54-8.42 ms —
first config where xserv wins decode outright. GSM8K-100: 96% (dense
FP8: 91%). llama TP=1 (2.9 ms) remains ahead: next levers are decode
CUDA graphs, non-expert quantization, sparse prefill (docs/20).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Weight-only 4-bit for the gpt-oss MoE experts: weights stored MXFP4 (E2M1 +
per-32-element UE8M0 block scale, tools/quantize_mxfp4.py), a fused kernel reads
the 4-bit weights and dequantizes on-chip to BF16. Decode (M=1) uses a fused
dequant-GEMV (batched_gemv_mxfp4) with shared-memory activation tiling; prefill
(M>1) dequantizes to BF16 then reuses the BF16 batched GEMM. MXFP4 is detected
by the scale tensor's rank (3-D [E,N,K/32]) vs FP8's 1-D [E].
Verified on dash5 (gpt-oss-20b, TP=2, 5090): byte-identical greedy tokens to
FP8/BF16, smallest footprint (13 GB vs 22 GB FP8, 39 GB BF16) — fits one 32 GB
5090 with room for KV cache.
NOT a decode speedup: the hand-written W4A16 GEMV (no tensor cores) is less
efficient than cuBLASLt's FP8 tensor-core GEMM, so even at half the weight bytes
decode is 17.0 ms vs FP8 13.5 ms (faster than BF16 18.8 ms); prefill regresses
(350 vs 134 ms, dequant fallback). Committed as a correct memory-optimization
foundation. Beating FP8 on speed needs FP4 tensor cores (W4A4, cuBLASLt
block-scaled MXFP4) or a Marlin-class kernel; see
docs/benchmarks/mxfp4-and-llama-decode.md.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Replace the W8A16 dequant→BF16-GEMM path with native FP8×FP8→BF16 GEMM
using cuBLASLt on Blackwell (RTX 5090). Both weights (static FP8 E4M3)
and activations (dynamically quantized per-row) are processed directly
on FP8 tensor cores.
Key implementation details:
- cuBLASLt on Blackwell requires transA=T for FP8, so expert weights
are transposed during model loading ([E,K,N] → [E,N,K])
- Per-row activation quantization kernel (absmax/448 → FP8 E4M3)
- Post-GEMM row-wise rescaling recovers per-token precision
- Per-expert loop (not batched) due to cuBLASLt FP8 scale constraints
The same FP8 quantized model files work — no re-quantization needed.
Activation quantization happens dynamically at inference time.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Close the <think> block when EOS or max_tokens interrupts an analysis
channel, and flush stdout after each analysis token so --think streams
smoothly instead of dumping in buffer-sized chunks.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The gpt-oss harmony `analysis` channel is the model's reasoning, analogous
to Qwen3's <think>. With --think, wrap it in a `<think>\n…\n</think>\n\n`
block (gray when color is on, like Qwen3) and then print the final-channel
answer; without --think, suppress the analysis and show only the answer.
Replaces the previous color-gated behavior (analysis shown gray only on a
TTY, with no markers). Analysis is still excluded from the multi-turn
history (answer_ids), so re-prefill drops CoT as before.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Re-render the whole conversation each turn and re-prefill into a freshly
cleared slot, with past assistant messages rendered as completed `final`
channels (analysis dropped, terminated with <|end|> not the <|return|>
stop token) — matching the model's training format and the server's
builder. The previous incremental cache kept every turn's chain-of-thought
plus <|return|> in context, which is out of distribution for harmony
multi-turn. The generator now returns the final-channel text to feed back
as history. Qwen3 keeps the incremental cache (its ChatML format is
unaffected); reset_slot factors out the free+re-register.
NOTE: this corrects the multi-turn *format* but does NOT cure the
long-context collapse on some inputs. That is a forward-pass numerical bug
(NaN / degenerate logits) reproducible in clean bench-gpt-oss independent
of the chat layer — the collapse token is vocab_size-1 (201087), the
all-NaN argmax tie-break. Tracked separately.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The hand-rolled gpt-oss system message dropped the canonical harmony
structure (identity / knowledge cutoff / current date / Reasoning level),
putting the model out of distribution — greedy decoding then flipped into
garbage or analysis loops on ~half of single-turn requests. Emit the
canonical system message (matching the model's chat_template.jinja
build_system_message macro) with Reasoning: low, plus a today_ymd() date
helper.
Also:
- Default the repetition penalty to off (1.0). Penalizing the harmony
control tokens (<|channel|>/<|message|>/<|start|>) that must repeat to
open the final channel made gpt-oss stop right after analysis, emitting
nothing.
- Suppress the literal "assistant" role header emitted between the
analysis and final channels (only print in the final channel, moe only;
non-moe Qwen3 stays in Normal and prints as before).
Verified on dash5 (TP=2): single-turn "capital of France" is now stable
across runs with a clean final answer; Qwen3 chat unaffected.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Let the model generate its own <|channel|> routing instead of forcing
<|channel|>final<|message|> — matches the GGUF chat template behavior.
- State machine tracks harmony channels: analysis channel rendered gray,
final channel printed normally, <|end|> stops on final channel only.
- Add repetition penalty (default 1.3 for MoE, 1.0 for Qwen) with 512
token window to prevent greedy decode loops. Configurable via
XSERV_REP_PENALTY and XSERV_REP_WINDOW env vars.
- Fix Length path: use <|end|> instead of <|im_end|> for gpt-oss to
avoid poisoning the KV cache with garbage tokens on truncation.
- Server api.rs: append <|channel|>final<|message|> to the hardcoded
gpt-oss prompt (server expects to post-process the JSON output).
- Add startswith filter to minijinja for harmony template compatibility.
Known issue: gpt-oss multi-turn NaN when total context exceeds ~256
tokens — likely a flash_attention_sinks kernel bug with sliding window
layers at large kv_len + small q_len. Single-turn and short multi-turn
conversations work correctly.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The gpt-oss harmony format generates internal control tokens
(<|channel|>, <|start|>, <|end|>, <|message|>) that should not appear
in the user-facing output. Additionally, <|end|> marks the end of a
response segment but was not in the model's EOS list, causing the
model to self-prompt into analysis channels and loop.
Fix: treat <|end|> as a stop token, skip all harmony special tokens
from the output stream.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add ChatModel enum dispatching between Qwen3 and GptOss based on
config.is_moe(), following the TP engine pattern.
- Add --tp N flag for tensor-parallel inference (required for 39GB
gpt-oss-20b which doesn't fit on a single 32GB GPU).
- Add gpt-oss harmony chat template with channel/message format.
- Replace hardcoded is_stop_token() with tokenizer.is_eos() for
multi-model EOS support.
- Restore gpt-oss hardcoded prompt template in server api.rs, lost
during the Jinja template refactor.
- Fix GEMV race condition: the K-split kernel zeroed the FP32
accumulator inside the kernel (block k=0) while other blocks
atomicAdd'd concurrently. Pre-zero with cudaMemsetAsync instead.
- Update benchmark docs with post-fix results.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the per-token CPU-routed MoE forward with an all-GPU path:
1. moe_topk_softmax: GPU top-k + softmax (was CPU sort + softmax)
2. moe_replicate: broadcast input to all local experts
3. cublasGemmStridedBatchedEx: batched expert matmul (was per-expert cuBLAS)
4. moe_weighted_sum: FP32-accumulated weighted sum on GPU (was GPU→CPU→F32→BF16→GPU)
Expert weights stored as contiguous 3D tensors for strided batched GEMM.
Zero CPU↔GPU transfers per MoE layer (was ~40 per token per layer).
Also: configurable geglu_alpha, LayerNorm bias auto-detect, unused-weight
diagnostic at load time.
GSM8K 30-problem: 11/30 → 23/30 (76.7%) vs llama.cpp 30/30 (100%).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add --prompt to override the fixed prompt, and two teacher-forced
diagnostics: --forced runs prefill over prompt+oracle ids and reports
per-position top-1 agreement; --forced-decode walks the oracle trajectory
through the decode path with per-position agreement bucketed by position,
to localize long-context decode divergence from the reference.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Make argmax skip NaN logits (warn once) instead of panicking the engine
thread on a single NaN. Add sample_greedy_penalized() applying an
HF-style repetition penalty over recent ids on the greedy path, to break
greedy repetition loops on reasoning models without touching the forward
pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add flash_attention_sinks_bf16 prefill kernel that folds the per-head
attention sink into the softmax denominator (exactly as the decode sink
kernel) and supports an optional sliding-window mask matching HF gpt-oss.
Wire it through xserv-kernels (flash_attention_sinks) and use it in
GptOss prefill, replacing the post-hoc sink approximation for an exact
match against the reference math.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The custom launch_gemv_bf16 kernel produces NaN when output dimension N
is small (e.g. N=32 for the MoE router). Fall back to cuBLAS GemmEx for
N < 256. Also removes the padding workaround in gpt_oss MoE forward.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Adds --pp N for layer-wise pipeline parallelism via NCCL P2P send/recv.
Each stage holds layers [s*L, (s+1)*L), stage 0 owns embedding, last
stage owns norm/lm_head. v1 serial (one request at a time) — correctness
+ per-GPU memory savings (~1/N). Refactors model to unfused QKV/gate_up
projections and removes unused kernels (argmax, reshape_and_cache).
Weight fusion at load time:
- q/k/v_proj → single qkv_proj_wt, GEMV once then narrow() to split
- gate/up_proj → single gate_up_proj_wt, same pattern
- Reduces GEMV calls from 7 to 4 per layer (36 layers → 108 fewer launches)
Batched decode refactor (forward_decode_paged):
- Per-head RMSNorm: reshape to [B*H, D], one rmsnorm call
- Batched RoPE: one call for all sequences
- Batched KV scatter: one reshape_and_cache kernel per layer
- Eliminates the per-sequence loop entirely
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the Rust cudaMemcpy loop in append_tokens() with the new
reshape_and_cache kernel. Add append_tokens_batched() for the decode
path using the batched variant.
Fix: use data_ptr() instead of storage().gpu_buffer().as_ptr() so that
tensor offset is respected. The old code silently read from storage base
(element 0) instead of the tensor's logical start, which produced wrong
results when K/V tensors were narrow() views into a fused QKV buffer.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Layer-wise split: each stage loads only its contiguous layer range
[s*L, (s+1)*L); stage 0 keeps embed_tokens, the last stage keeps
norm/lm_head (others get a 1x1 placeholder). Heads are NOT split
(PP is orthogonal to TP). Adds embed/head and forward_layers_prefill/
forward_layers_decode that take and return the [tokens, hidden] hidden
state; per-stage PagedKVCache is indexed by local layer id.
sampling: derive Clone on SamplingParams (carried in the PP command enum).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Cooked-mode read_line() left line editing to the terminal, so Backspace on a
multi-byte 汉字/かな/한글 deleted a byte (or behaved inconsistently across TTYs).
Replace with a raw-mode reader (libc termios): Backspace pops a whole char,
multi-byte input is reassembled from its continuation bytes, and a full-line
redraw renders double-width glyphs correctly. Non-TTY input falls back to a
plain read; raw mode is restored after each line. libc is already a locked
transitive dep, so this builds offline.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
from_weights_tp shards each rank's weights (column-split q/k/v/gate/up,
row-split o/down; replicate norms/embed/lm_head) and the paged forward uses
local head counts + AllReduces after o_proj and down_proj. PagedKVCache::new_tp
sizes the pool for the rank's local KV heads (KV is sharded too). TP=1 is the
identity path. New bench-tp binary runs E2E multi-GPU generation per TP degree.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- paged_kv_cache: new block-paged KV cache; adds a pinned-host swap pool with
a second BlockAllocator, per-sequence Location {Gpu,Cpu}, and lossless
swap_out/swap_in (block-granular D2H/H2D) for vLLM-style preemption.
bytes_per_block helper exposes per-block cost for VRAM-based sizing.
- decode_graph: CUDA-graph decode path.
- qwen3/gpt2/kv_cache: paged prefill/decode forward + related updates.
- tokenizer/bins: BPE updates, new xserv-chat CLI, bench-qwen3 tweaks.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Three performance optimizations targeting decode throughput:
1. Decode Attention Kernel (csrc/attention/flash_attention.cu):
- Specialized kernel for Q_len=1 (decode step)
- 256 threads parallelize across KV sequence dimension
- Online softmax with block-level warp-shuffle reduction
- Replaces FA2 kernel which wasted 63/64 threads for decode
- flash_attention() auto-dispatches when q_len==1
2. Fused SiLU×Mul (csrc/activation/activations.cu):
- Single kernel: out = silu(gate) * up
- Saves 1 HBM read + 1 HBM write per FFN layer (N elements)
- Eliminates intermediate tensor allocation
3. Fused Add+RMSNorm (csrc/normalization/rmsnorm.cu):
- Single kernel: (normed, sum) = (rmsnorm(x+residual), x+residual)
- Saves 1 full HBM round-trip per attention block
- Eliminates separate add + rmsnorm kernel pair
Performance analysis:
- At current short sequences (max 79 tokens), these optimizations provide
marginal benefit because the bottleneck is cuBLAS GEMV overhead:
252 weight matrix reads × ~32MB each = 15.5 GB per decode step.
Theoretical minimum at 1.79 TB/s = 8.7ms, actual ~78ms (9x gap).
- The fused kernels and decode attention will show larger gains at
longer sequences where attention and element-wise ops dominate.
- Next optimization target: CUDA Graphs to eliminate kernel launch
overhead, or custom GEMV kernels to replace cuBLAS for M=1.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- GpuKVCache: pre-allocated GPU buffers, D2D copy append at offset
- Per-head strided layout [num_kv_heads, max_seq_len, head_dim]
- Fixed critical bug: seq_len must advance AFTER all layers write
(not inside the loop per-layer)
- GpuBuffer::copy_from_device_at for offset-based D2D copy
- Tensor::from_storage constructor for wrapping raw GPU buffers
- Exported Storage and Dims from xserv-tensor
Correctness: GPU KV cache vs CPU KV cache = 50/50 bit-identical
Performance: ~neutral (KV cache was never the main bottleneck —
reshape/merge/transpose CPU round-trips dominate for Qwen3-8B)
TTFT: 122ms, TBT: 142ms, 7.0 tok/s (marginal change from 7.3)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Kernel additions:
- add_f32/bf16, mul_f32/bf16 CUDA kernels (element-wise, on GPU)
- Refactored activation.rs with dispatch_unary/dispatch_binary helpers
- Qwen3 and GPT-2 now use GPU add/mul instead of CPU round-trips
GPT-2 add_bias also moved to GPU (broadcast via tile + GPU add)
BF16 precision analysis (docs/benchmarks/phase10-qwen3.md):
- Root cause: separate attention kernels materialize BF16 intermediates
(QK^T→BF16→scale→BF16→mask→BF16→softmax→BF16 vs HF's fused FP32 path)
- HF itself SDPA vs Eager also differs by ~0.125 logit
- xserv vs HF: ~1-2 logit systematic offset, but same top-1 in 84% cases
- Industry standard for BF16: top-5 overlap (we achieve 100%)
- Fix path: Flash Attention (Phase 14) to fuse attention in FP32
Performance: TTFT 138→119ms, TBT 144→137ms (GPU ops faster than CPU)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Qwen3 model (qwen3.rs):
- RMSNorm + QK normalization (per-head q_norm/k_norm)
- GQA: 32 Q heads, 8 KV heads, repeat_kv for attention
- SwiGLU FFN: gate_proj → SiLU → * up_proj → down_proj
- RoPE with transpose for [1,H,S,D] ↔ [S,H,D] layout
- BF16 forward pass, [out,in] weight layout via linear_t
- No attention bias (attention_bias=false)
Tokenizer fixes:
- Fixed unicode_to_byte: shifted bytes now use correct inverse lookup table
- MergeEntry supports both string and array formats
- Both GPT-2 and Qwen3 tokenizers work correctly (English + Chinese)
KVCache refactored:
- Dtype-agnostic: stores raw bytes per-head, works for F32 and BF16
- append_kv_tensor/get_kv_tensors use Tensor directly
CLI updated:
- Auto-detects model type from config.json (gpt2 vs qwen3)
- Supports both GPT-2 (F32) and Qwen3 (BF16)
Verified: Qwen3-8B generates coherent English and Chinese on single RTX 5090.
61/61 tests pass, GPT-2 performance no regression.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Phase 6 — Model Loading (xserv-model):
- safetensors parser with single/sharded file support
- ModelConfig with dual naming (GPT-2 n_embd/n_head + modern HF naming)
- Weight loading flow: safetensors → mmap → CPU Tensor → GPU
Phase 7 — BPE Tokenizer (xserv-tokenizer):
- Full BPE encode/decode from tokenizer.json
- GPT-2 byte-to-unicode mapping (printable ASCII identity + shifted bytes)
- Pre-tokenization regex, special token handling
- Chat template support structure
Phase 8 — GPT-2 Complete Inference:
- GPT-2 model definition: wte, wpe, 12 transformer blocks, ln_f
- Forward pass: embedding → (LayerNorm → MHA → residual → LayerNorm → MLP → residual) × 12 → LN → logits
- QKV split with correct [batch, heads, seq, dim] layout (fixed reshape bug)
- Greedy sampling from last-position logits
- Interactive CLI: xserv-cli <model-dir> [--max-tokens N]
Verified: GPT-2 124M generates coherent English text on RTX 5090.
"The future of AI is uncertain. The future of AI is uncertain..."
"Once upon a time, the world was a place of great beauty..."
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
