flash_attention_sinks_bf16_kernel skipped only fully-future KV tiles (the
causal `continue`); an early tile entirely outside the sliding window was
still processed with every key masked to -inf, so row_max == -INFINITY.
Folding that into the online softmax computed expf(-inf - (-inf)) = NaN,
and the next valid tile's 0*NaN correction then poisoned the whole row.
Result: the gpt-oss prefill produced all-NaN logits for any query whose
sliding window (128) starts past the first KV tile — i.e. at longer
context — collapsing generation into a single repeated token (argmax of
all-NaN logits: vocab_size-1 in bench, token 0 "!" in the chat). This was
the residual multi-turn/long-context collapse.
Fix: skip a fully-masked tile (row_max == -INFINITY) — it contributes
nothing to the softmax. The decode kernel already guards
local_max == -INFINITY, so it was unaffected.
Verified on dash5 (TP=2): the prefill that previously went all-NaN now
produces clean logits; multi-turn gpt-oss chat (e.g. a haiku after a long
prior answer) completes correctly instead of emitting "!!!!".
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>
build_prompt_gpt_oss (the hardcoded builder used when a gpt-oss model
ships no Jinja chat template) emitted the same malformed "You are a
helpful assistant." system message that destabilized the CLI. Replace it
with the canonical harmony system message (identity / knowledge cutoff /
current date via strftime_now / Reasoning: low / channels), matching the
chat_template.jinja build_system_message macro and the xserv-chat fix.
Dormant for gpt-oss-20b (it ships a Jinja template, so the template path
runs), but correct now for any gpt-oss model without one.
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>
Load the model's chat_template.jinja (or tokenizer_config.json
chat_template field) at startup and render it with minijinja instead of
hardcoded per-model prompt builders.
Custom Jinja functions: strftime_now (date formatting), raise_exception
(template validation errors). Falls back to Qwen3 ChatML template if
no Jinja template is found.
Removes the hardcoded build_prompt_gpt_oss() — the model's own template
now drives prompt formatting, matching llama.cpp's behavior exactly.
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>
Parse the model's `pre_tokenizer` section to extract its Split regex
instead of hardcoding the GPT-2 pattern. The gpt-oss-20b model uses
a GPT-4-style regex that produces different word boundaries, causing a
1-token prompt mismatch vs HuggingFace (136 → 135 tokens, now aligned).
Unsupported lookahead `(?!\S)` is stripped — it only affects trailing
whitespace edge cases. Falls back to the old GPT-2/Qwen heuristic if
the model regex fails to compile or is absent.
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>
Route caller-supplied system messages into a harmony 'developer'
instructions block (<|start|>developer<|message|># Instructions...),
keeping the fixed system/meta block for the channel declaration. Harmony
puts user instructions on the developer role, not system.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Use tokenizer.is_eos() (multi-eos) for generation termination in both PP
and TP engines instead of a single eos id, so gpt-oss stops on <|return|>
/<|call|>/<|endoftext|>.
In the TP engine, optionally apply a repetition penalty on the greedy
decode path (XSERV_REP_PENALTY>1 over XSERV_REP_WINDOW recent tokens; off
by default) to break greedy repetition loops.
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>
Track an ordered eos_token_ids list (not just one id) and add is_eos().
gpt-oss/harmony ends the assistant turn on <|return|> and also treats
<|call|> and <|endoftext|> as terminators (generation_config.json
eos_token_id = [200002, 199999, 200012]); single-eos families are
unchanged.
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 gpt-oss model requires a specific prompt format with <|start|>,
<|message|>, <|end|>, <|channel|> tokens. Without this, the model
produces degenerate output. Auto-detected via config.model_type.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- tp_engine.rs: TpModel enum dispatches between Qwen3 and GptOss based on
config.is_moe(). Server auto-detects model type on startup.
- tools/run_gpt_oss_bench.sh: one-click benchmark comparing xserv (TP=2)
vs llama.cpp (BF16 GGUF) on GSM8K quality + speed
Co-Authored-By: Claude Opus 4.6 (1M context) <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).
When all active sequences use temperature=0, run argmax on the GPU and
only D2H the token ids (~B×4 bytes) instead of the full [B, vocab_size]
BF16 logits (~1.2 MB at B=4, Qwen3 vocab=152K). Mixed-sampling batches
fall back to the existing CPU path.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
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>
Three new CUDA kernels and one rewrite:
- reshape_and_cache: scatter K/V into paged pool in a single kernel per
layer, replacing the Rust-side per-token per-head cudaMemcpy loop.
Includes both single-sequence (prefill) and batched (decode) variants.
- argmax: GPU-side BF16 argmax with warp-shuffle reduction. Greedy
decode now only D2H-transfers B×4 bytes (token ids) instead of the
full [B, vocab] logits tensor.
- GEMV rewrite: fused zero-init inside the K-split kernel eliminates
the cudaMemsetAsync call, reducing launches from 3 to 2 per GEMV.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Exposes the caching allocator's trim() through a public free function.
Called after weight fusion during model loading to free temporary buffers
that would otherwise sit in the pool and cause OOM.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
narrow(dim, start, len) creates a zero-copy slice along any dimension.
is_contiguous() now ignores stride mismatches on dimensions of size 1,
since those dimensions are never stepped. This avoids unnecessary GPU
strided copies when slicing fused projection outputs at batch=1.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
pp_engine::run_pp: stage-0 coordinator (scheduler/tokenizer/sampling +
stop logic) on the calling thread, worker stage threads for 1..P. Each
step the coordinator embeds + runs its layers, then the hidden state is
handed stage->stage over NCCL P2P; the last stage samples and returns
the token to stage 0 over an in-process channel. v1 is serial (one
request, one token/step) — correctness first; throughput via microbatch
overlap is future work.
main: wire --pp N (mutually exclusive with --tp).
Co-Authored-By: Claude Opus 4.8 <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>
Add ncclSend/ncclRecv FFI and a PpContext that initializes a NCCL
communicator across P pipeline stages and hands the hidden state to
neighbour stages on the null stream. Mirrors TpContext; the collective
differs (point-to-point hand-off vs in-layer AllReduce).
tests/sendrecv.rs: 2-GPU stage0->stage1 send/recv smoke test.
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>
AIME 2025 + GSM8K at TP=1/2/4. Quality on par across engines/TP. Opposite
perf scaling: xserv TPOT improves with TP (21->17->15ms) while llama.cpp
row-split regresses over PCIe (10->19->20ms), crossing over so xserv is faster
at TP=4. Includes the clean same-path bench-tp scaling (58/76/86 tok/s).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
runner/servers gain --tp (xserv --tp N; llama.cpp --split-mode row) and
--llama-devices so llama can run on a disjoint GPU group. run_tp_parallel.sh
runs xserv (GPU 0..N-1) and llama.cpp (GPU 4..4+N-1) concurrently per TP,
matching the box's 0-3 / 4-7 PHB groups. summarize_tp.py tabulates the sweep.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
tp_engine: rank-0 coordinator owns the scheduler and broadcasts per-token
commands (Register/Prefill/Decode/Free) to worker rank threads; the sampled
token always comes from rank 0, so it's correct for greedy and stochastic
sampling. Serial single-request path (sufficient for the quality benchmark).
--tp N selects it; TP=1 keeps the existing single-GPU Engine unchanged.
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>
New xserv-distributed crate: hand-written NCCL FFI, TpContext (one rank per
thread, bound to one GPU), and in-place BF16 AllReduce on the null stream so
it orders naturally with the model's kernels. 2-GPU AllReduce test included.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Megatron-style TP for Qwen3 on the 8x5090 (no-NVLink, PCIe) box: column/row
split per layer, 2 AllReduces/layer, multi-thread one-rank-per-GPU model,
NCCL, sharded weights, and the incremental implementation + verification plan.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Project intro, architecture, build, basic usage (HTTP server / CLI / bench),
and the llama.cpp comparison workflow.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Re-ran the full comparison at --max-seq-len 8192 now that xserv handles it:
- OOM finding resolved — pool sized to available VRAM + vLLM-style host swap;
8192 runs with 0 swap events (swap is the overload safety net).
- Quality at parity with equal context: AIME 20.0% vs 20.0%, GSM8K 98% vs 96%.
- Speed unchanged relative to llama.cpp (~0.42-0.60x); TPOT is bandwidth-bound.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Fixes the paged-KV OOM at large --max-seq-len and adds elastic memory:
- Size the GPU block pool to available VRAM (cudaMemGetInfo) instead of the
worst-case blocks_per_seq * max_batch * 2 reservation, which OOM'd at 8192.
- Scheduler tracks waiting/running/swapped sets: block-aware admission,
swap-in of resumable sequences when blocks free, and preemption of the
newest running sequence to host when the pool can't cover a decode step.
- --swap-space-gb (default 8) sizes the pinned host swap pool;
XSERV_MAX_KV_BLOCKS forces a small pool to exercise swapping.
- api: poison-tolerant lock + clean 503 when the engine thread is gone,
instead of cascading mutex-poison panics.
Verified on RTX 5090: serves at --max-seq-len 8192 (previously OOM), and a
forced 40-block pool drives 48 lossless swap-out/swap-in cycles under
concurrency with coherent output.
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>
CUDA layer for the paged-KV + swap work:
- csrc: new paged_attention.cu plus updates across attention/gemm/norm/
activation/embedding/reduce kernels and common.cuh.
- xserv-kernels: new dispatch module and kernel-binding updates.
- xserv-cuda: cudaMallocHost/FreeHost bindings + PinnedBuffer (host swap
pool backing) and offset-aware D2H/H2D copies used to move KV blocks
between the GPU pool and pinned host memory.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Record what the new baseline adds (llama.cpp pinned b9371, same BF16 weights,
AIME 2025 + GSM8K) and the measured results: performance (xserv ~0.45-0.61x
llama.cpp throughput) and quality parity (GSM8K 94% vs 96%, AIME 23.3% vs 20%
after the context fix), plus the findings the bench surfaced.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
llama.cpp divides total -c across --parallel slots, so -c 4096 --parallel 4
gave each request only 1024 tokens — truncating long AIME generations before
the boxed answer and making xserv look artificially better (20% vs 3.3%).
Set total -c = max_seq_len * n_parallel so per-slot context equals xserv's
per-sequence max_seq_len. Also drop --log-disable; its startup log reports the
per-slot n_ctx that catches exactly this misconfiguration.
After the fix, AIME is at parity (xserv 23.3% vs llama.cpp 20.0%), matching the
GSM8K parity and confirming the gap was a config artifact, not engine quality.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Refinements from end-to-end bring-up on the GPU host:
- Run each system start→suites→stop in sequence. Two BF16 8B models don't
co-reside on one 32GB GPU, and a resident idle engine would distort the
other's latency/throughput.
- Match generation mode: xserv hardcodes Qwen3 thinking off, so send
chat_template_kwargs={enable_thinking:false} to llama.cpp via a per-endpoint
extra_body. --enable-thinking opts back into thinking mode.
- Add tools/__init__.py so `python3 -m tools.bench.runner` resolves our package
instead of a site-packages `tools` (nvfuser ships one that shadowed it).
- Document offline-GPU-host workflow, thinking-match, and the xserv 8192 OOM
finding that the bench surfaced.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Vendor llama.cpp as a submodule pinned to b9371 and add a one-click
benchmark driver that compares xserv against it on identical workloads:
- setup-llama-cpp.sh: network-optional CUDA build (SM120); convert-to-gguf.sh
converts the same safetensors to BF16 GGUF for an apples-to-apples baseline.
- tools/bench/: black-box OpenAI-API driver measuring TTFT/TPOT/throughput
(single-stream + concurrent) and response quality on AIME 2025 + GSM8K.
- fetch_datasets.py pulls datasets to local JSON (GPU host has no network);
task loaders prefer the local JSON.
- sync-and-build.sh: `bench` subcommand transfers source + datasets to the
GPU host via tar-over-ssh (no rsync there), builds, and runs the suite.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
