Add files using upload-large-folder tool

.gitattributes

@@ -58,3 +58,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+profiles/perf_compile.nsys-rep filter=lfs diff=lfs merge=lfs -text
+profiles/perf.sqlite filter=lfs diff=lfs merge=lfs -text
+profiles/perf.nsys-rep filter=lfs diff=lfs merge=lfs -text
+profiles/perf_compile.sqlite filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,82 @@
+---
+license: cc-by-4.0
+tags:
+  - profiling
+  - nsight-systems
+  - pytorch
+  - video-generation
+  - sequence-parallelism
+pretty_name: fastvideo nsys profile — Wan2.1-T2V-1.3B on 4× L40S (PCIe Gen4, no NVLink)
+size_categories:
+  - 1B<n<10B
+---
+
+# fastvideo / Wan2.1-T2V-1.3B — nsys profiles on 4× L40S (no NVLink)
+
+This dataset hosts NVIDIA Nsight Systems profiles of [fastvideo](https://github.com/hao-ai-lab/FastVideo) running `Wan-AI/Wan2.1-T2V-1.3B-Diffusers` text-to-video inference on a 4× L40S node (PCIe Gen4 x16, no NVLink), plus the static plots and measurement summary used to discuss the trace.
+
+The trace shows a **bandwidth-bound, same-stream-serialized** workload: compute kernels and NCCL `SendRecv` collectives share a single CUDA stream, so there is no compute/comm overlap to exploit — every NCCL byte directly inflates forward time.
+
+## Contents
+
+```
+profiles/
+  perf.nsys-rep            (605 MB) — non-compile baseline
+  perf.sqlite              (2.1 GB) — non-compile baseline, exported
+  perf_compile.nsys-rep    (506 MB) — torch.compile enabled
+  perf_compile.sqlite      (1.8 GB) — torch.compile enabled, exported
+
+plots/
+  1_iter_duration_over_time.png    — per-iter latency vs wall time (warmup shaded)
+  2_iter_duration_histogram.png    — steady-state distribution (σ = 21 ms / 0.66 %)
+  3_stream7_one_iter.png           — stream-7 timeline showing compute / NCCL alternation
+  4_per_device_walltime.png        — 47 % compute / 34 % NCCL / 18 % idle per GPU
+  5_nccl_kernel_duration_hist.png  — bimodal at 13 ms / 33 ms (two AllToAll4D sizes)
+  README.md                        — mentor-question Q&A walkthrough
+
+analysis_measurements.md           — raw §1–§9 measurements from perf_compile.sqlite
+```
+
+## Headline numbers (perf_compile.sqlite, full 677 s span, device 0)
+
+| Metric | Value |
+| --- | --- |
+| `compute_only_ms` | 311 516 ms (46.0 %) |
+| `nccl_only_ms` | 271 564 ms (40.1 %) |
+| `idle_ms` | 94 265 ms (13.9 %) |
+| `overlap_ms` | **0.02 ms (0.0 %)** |
+| `same_stream_compute_pct` | 100 % on stream 7 |
+| `same_stream_nccl_pct` | 100 % on stream 7 |
+| Total NCCL payload | 8 936 GiB |
+| NCCL message p50 / p99 | 13.18 / 39.55 MiB |
+| Slowest iter | iter 311 / 3 087 ms (compile); iter 321 / 3 292 ms (no-compile) |
+| Steady-state σ | 17.7 ms (compile) / 21.2 ms (no-compile) |
+
+## How to reproduce the measurements
+
+```bash
+pip install nsys-ai
+
+# Health + topology + bandwidth ceiling
+nsys-ai analyze profile_health_manifest    perf_compile.sqlite
+nsys-ai analyze overlap_breakdown          perf_compile.sqlite
+nsys-ai analyze nccl_payload_breakdown     perf_compile.sqlite
+nsys-ai analyze iteration_timing           perf_compile.sqlite
+nsys-ai analyze nvtx_layer_breakdown       perf_compile.sqlite
+
+# Or just open the timeline in the browser
+nsys-ai timeline-web perf_compile.sqlite
+```
+
+The plots in `plots/` were generated from `perf.sqlite` via a matplotlib script that queries the parquet cache `nsys-ai` builds on first load (no LLM, no API).
+
+## Hardware / software
+
+- 4× NVIDIA L40S, PCIe Gen4 x16, no NVLink (intra-node bus only)
+- NCCL aggregate bus BW ≈ 24 GB/s (2 channels × 12 GB/s) measured during the same session
+- PyTorch + fastvideo, Ulysses-style sequence parallelism (`all_to_all_4D` scatter=2/gather=1 pre-attn, scatter=1/gather=2 post-attn)
+- Two passes per denoising step (CFG cond + uncond) — 90 denoising steps × 2 = 180 real forwards
+
+## License
+
+Profile traces and derived artifacts are released under CC-BY-4.0.

analysis_measurements.md

@@ -0,0 +1,259 @@
+## 1. Profile-level row counts (direct SQL)
+
+```
+NVTX_EVENTS                  10,457,403
+CUPTI_ACTIVITY_KIND_KERNEL    1,079,198
+CUPTI_ACTIVITY_KIND_RUNTIME   7,165,943
+CUPTI_ACTIVITY_KIND_MEMCPY      663,807
+CUPTI_ACTIVITY_KIND_SYNCHRONIZATION 2,737,429
+CUPTI_ACTIVITY_KIND_OVERHEAD    188,445
+NVTX_PAYLOAD_SCHEMAS                 24 rows (6 distinct schemaId)
+NVTX_PAYLOAD_SCHEMA_ENTRIES          68 rows
+NVTX_EVENTS.binaryData       346,400 non-null
+StringIds                     7,989,142
+```
+
+Kernel-table span: `MIN(start) → MAX(end)` = 677,543 ms (677.5 s wall).
+
+Distinct deviceIds in `CUPTI_ACTIVITY_KIND_KERNEL`: `[0, 1, 2, 3]`.
+
+---
+
+## 2. `profile_health_manifest`
+
+```json
+gpu: "unknown"
+fingerprint: {
+  framework: "DeepSpeed",
+  distributed: true,           // PR #127 fallback fired (kernel-table distinct deviceIds > 1)
+  multi_node: false,
+  nic_summary: "",
+  precision_notes: []
+}
+data_quality.auto_trim: {
+  applied: true,
+  trim_start_ns: 394,791,526,439,
+  trim_end_ns:   414,791,526,439,
+  window_ms:      20,000.0,
+  profile_full_span_ms: 677,543.7
+}
+data_quality.overhead_pct: 28.6
+nvtx: {
+  has_nvtx: true,
+  iteration_count: 16,             // manifest's own detection
+  median_iter_ms: 2.3,
+  slowest_iter_ms: 2,944.2,
+  top_regions: [
+    { name: "aten::to,        op_id=1298527", total_ms: 2,310.5, count: 2 },
+    { name: "aten::_to_copy,  op_id=1298528", total_ms: 2,310.5, count: 1 },
+    { name: "aten::copy_,     op_id=1298530", total_ms: 2,310.5, count: 1 },
+    { name: "aten::to,        op_id=1298528", total_ms: 2,310.3, count: 1 },
+    { name: "aten::_to_copy,  op_id=1298529", total_ms: 2,310.3, count: 1 },
+  ]
+}
+suspected_bottleneck: "High CPU Synchronization Blocking (37.0% of span)"
+```
+
+**All fields above describe the auto-trimmed 20-second window, not the full
+677.5-second profile.** Numbers below for `overlap_breakdown`, `sync_cost_analysis`,
+and `nccl_payload_breakdown` cover the full profile.
+
+---
+
+## 3. `overlap_breakdown` (device 0, full profile)
+
+```
+total_ms                  677,344.6
+compute_only_ms           311,515.7  (46.0 %)
+nccl_only_ms              271,563.8  (40.1 %)
+overlap_ms                      0.02 ( 0.0 %)
+idle_ms                    94,265.1  (13.9 %)
+sync_ms                   214,805.6
+compute_kernels           258,818
+nccl_kernels               10,985
+span_start_ns          39,565,504,542
+span_end_ns           716,910,057,111
+
+same_stream_diagnosis    ["7"]
+same_stream_compute_pct  100.0          // PR #127 addition
+same_stream_nccl_pct     100.0          // PR #127 addition
+present_devices          [0, 1, 2, 3]   // PR #127 addition
+device_id                 0
+```
+
+### 3a. Per-device kernel/stream layout (direct SQL across all 4 ranks)
+
+```
+dev=0  stream=7   compute n=258,818  total=311,515.7 ms
+                  nccl    n= 10,985  total=271,563.8 ms
+dev=1  stream=17  compute n=258,816  total=315,914.0 ms
+                  nccl    n= 10,985  total=267,733.3 ms
+dev=2  stream=17  compute n=258,810  total=314,592.1 ms
+                  nccl    n= 10,985  total=268,910.4 ms
+dev=3  stream=17  compute n=258,814  total=314,679.5 ms
+                  nccl    n= 10,985  total=268,502.0 ms
+```
+
+---
+
+## 4. `sync_cost_analysis` (full profile)
+
+```
+profile_span_ms          677,543.7
+total_sync_wall_ms       214,805.6
+sync_density_pct              31.7
+
+sync_by_type_ms:
+  STREAM_SYNCHRONIZE     211,502.5
+  EVENT_SYNCHRONIZE        2,673.0
+  CONTEXT_SYNCHRONIZE        738.5
+  STREAM_WAIT_EVENT          157.1
+```
+
+---
+
+## 5. `memory_transfers` (full profile, summed across all devices)
+
+```
+copyKind=1  H2D    count=597,708   bytes=2,764,132 MB   total_ms=324,398.0
+copyKind=2  D2H    count=  5,639   bytes=  108,113 MB   total_ms= 70,941.1
+copyKind=8  D2D    count= 60,460   bytes=5,058,436 MB   total_ms= 13,293.0
+```
+
+### 5a. Per-device breakdown (direct SQL `GROUP BY deviceId, copyKind`)
+
+```
+dev=0  H2D  149,427 calls   659,021 MB     87,205.4 ms
+       D2H    1,412 calls    27,603 MB     18,047.6 ms
+       D2D   15,118 calls 1,206,183 MB      3,319.7 ms
+dev=1  H2D  149,427 calls   659,021 MB     87,485.3 ms
+       D2H    1,409 calls    25,167 MB     17,570.1 ms
+       D2D   15,115 calls 1,206,025 MB      3,325.1 ms
+dev=2  H2D  149,427 calls   659,021 MB     74,761.9 ms
+       D2H    1,409 calls    25,167 MB     17,666.5 ms
+       D2D   15,113 calls 1,205,920 MB      3,323.7 ms
+dev=3  H2D  149,427 calls   659,021 MB     74,945.4 ms
+       D2H    1,409 calls    25,167 MB     17,656.9 ms
+       D2D   15,114 calls 1,205,972 MB      3,324.5 ms
+```
+
+---
+
+## 6. `nccl_payload_breakdown` (PR #128, full profile)
+
+```
+total_payload_events      346,400
+message_carrying_events   346,388
+skipped_events                  0
+total_bytes_all     9,594,962,460,672    // 8,936.01 GiB
+distinct_schemas                6
+```
+
+| schema_id | category | calls | distinct_communicators | msg p50 | msg p99 | msg max | bytes total |
+|---|---|---:|---:|---:|---:|---:|---:|
+| 16777218 | init | 4 | 1 | — | — | — | — |
+| 16777220 | collective | 724 | 1 | 52.73 MiB | 52.73 MiB | 52.73 MiB | 37.08 GiB |
+| 16777225 | p2p | 172,832 | 1 | 13.18 MiB | 39.55 MiB | 39.55 MiB | 4,449.46 GiB |
+| 16777226 | p2p | 172,832 | 1 | 13.18 MiB | 39.55 MiB | 39.55 MiB | 4,449.46 GiB |
+| 16777227 | init | 4 | 1 | — | — | — | — |
+| 16777230 | group_marker | 4 | 1 | — | — | — | — |
+
+### 6a. Schema field layouts (from `NVTX_PAYLOAD_SCHEMA_ENTRIES`)
+
+```
+16777218 (init, 24B):       NCCL communicator ID, No. of ranks, Rank, CUDA device
+16777220 (collective, 16B): NCCL communicator ID, Message size [bytes]
+16777225 (p2p, 24B):        NCCL communicator ID, Message size [bytes], Peer rank
+16777226 (p2p, 24B):        NCCL communicator ID, Message size [bytes], Peer rank
+16777227 (init, 24B):       NCCL communicator ID, No. of ranks, Rank, CUDA device
+16777230 (group_marker, 8B): NCCL communicator ID
+```
+
+---
+
+## 7. `iteration_timing` (full profile)
+
+```
+total rows returned                984
+rows with duration_ms > 1,000      180
+median (over 180 rows)         2,933.3 ms
+mean   (over 180 rows)         2,936.7 ms
+sum    (over 180 rows)         528,610   ms  (=528.6 s)
+```
+
+### 7a. Top 5 iterations by `duration_ms`
+
+```
+iter=311  dur=3,087.4 ms  kernels=1,172  compute=3,071.8 ms  nccl_count=59  text=heuristic_step_311
+iter=316  dur=2,991.4 ms  kernels=1,234  compute=2,975.1 ms  nccl_count=61  text=heuristic_step_316
+iter=321  dur=2,985.8 ms  kernels=1,212  compute=2,971.9 ms  nccl_count=61  text=heuristic_step_321
+iter=364  dur=2,983.1 ms  kernels=1,234  compute=2,966.8 ms  nccl_count=61  text=heuristic_step_364
+iter=313  dur=2,981.9 ms  kernels=1,238  compute=2,965.5 ms  nccl_count=61  text=heuristic_step_313
+```
+
+`text` prefix `heuristic_step_*` indicates the labels are synthesized by
+`iteration_timing`'s kernel-gap detection heuristic, not present in the raw
+NVTX_EVENTS (verified by `SELECT COUNT(*) FROM NVTX_EVENTS WHERE text LIKE 'heuristic_step%'` = 0).
+
+### 7b. `iteration_detail` on iter=311
+
+```
+duration_ms       3,087.35
+gpu_start_ns      109,428,571,984
+gpu_end_ns        112,515,921,387
+kernel_count      1,172
+nccl_count        59
+compute_ms        3,071.77
+median_ms         0.79         // median of all 984 iteration_timing rows
+vs_median        +390,703.8 %
+top_kernels:
+  ncclDevKernel_SendRecv             1,616.75 ms × 58  (54.7 %)
+  flash_fwd                            920.48 ms × 58  (31.1 %)
+  cutlass_80_tensorop_bf16_s16816gemm  122.29 ms × 203 ( 4.1 %)
+  cutlass_80_tensorop_bf16_s16816gemm   56.86 ms ×  29 ( 1.9 %)
+  at::elementwise_kernel                56.77 ms ×  58 ( 1.9 %)
+```
+
+---
+
+## 8. `top_kernels` (full profile, top 10 by total_ms)
+
+```
+ncclDevKernel_SendRecv                                  total_ms=1,051,241.5  inv=43,208  avg=24.33  tc=false
+void flash::flash_fwd_kernel<...,(int)128,(int)128,(int)32...>  648,214.3  inv=43,200  avg=15.00  tc=true
+void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_256x128_32x3...>  95,138.0  inv=151,200  avg=0.63  tc=true
+void at::native::elementwise_kernel<(int)128, (int)2, ...>    81,499.8  inv=77,100   avg=1.06  tc=false
+sm86_xmma_fprop_implicit_gemm_tf32f32_tf32f32_f32_nhwckrsc_nchw_tilesize128x128x...  64,636.5  inv=7,200  avg=8.98  tc=true
+void at::native::elementwise_kernel                              45,290.6  inv=53,832   avg=0.84  tc=false
+void cutlass::Kernel2<cutlass_80_tensorop_bf16_s16816gemm_relu_bf16_128x256_32x3...>  44,807.9  inv=21,600  avg=2.07  tc=true
+void at::native::elementwise_kernel                              29,725.7  inv=109,464  avg=0.27  tc=false
+void cudnn::engines_precompiled::nchwToNhwcKernel                27,721.9  inv=16,752   avg=1.65  tc=false
+void at::native::vectorized_elementwise_kernel                   27,692.3  inv=29,124   avg=0.95  tc=false
+```
+
+`total_ms` values are summed across all 4 device IDs (skill currently does not
+expose per-device totals); per-rank averages = total ÷ 4.
+
+---
+
+## 9. `nvtx_layer_breakdown` (top 8 regions by total_gpu_ms)
+
+```
+detection_method  numbered_pattern
+layer_depth       1
+layer_count       673
+confidence        1.0
+
+total_gpu_ms  kernels  nccl%   tc%      nvtx_path (truncated to 65 chars)
+    4,645.6   24,473   10.7   100.0    stage::DenoisingStage > Torch-Compiled Region: 0/0, op_id=74092
+      685.0       51   91.9   100.0    stage::DenoisingStage > Torch-Compiled Region: 0/0, op_id=90955
+      600.7      192    0.0   100.0    stage::DenoisingStage > aten::nonzero, op_id=1728365
+      434.8       55   78.9   100.0    stage::DenoisingStage > Torch-Compiled Region: 0/0, op_id=90929
+      405.3      101   14.9   100.0    stage::DenoisingStage > aten::to, op_id=1613263
+      354.8      113   14.2   100.0    stage::DenoisingStage > Torch-Compiled Region: 0/0, op_id=908401
+      339.5      112   14.1   100.0    stage::DenoisingStage > Torch-Compiled Region: 0/0, op_id=907210
+      333.2      107   13.8    91.9    stage::DenoisingStage > Torch-Compiled Region: 0/0, op_id=1722730
+```
+
+---
+

plots/README.md

@@ -0,0 +1,37 @@
+# Mentor Q&A — fastvideo perf.sqlite (4× L40S, no NVLink)
+
+Profile: `/home/rich-wsl/fastvideo/profile_results/perf.sqlite` — 2.04 GiB, 778 s GPU span.
+Generated 2026-05-16 from nsys-ai cached parquets.
+
+## Q1. Can we see the nsys profile / graphs?
+
+Yes — these 5 PNGs are static views of the profile. The full interactive timeline is also available via `nsys-ai timeline-web` (browser-based) if you want to scrub around. For the SSH-tunnel-free version, these screenshots cover the headline findings.
+
+## Q2. Which iter is the slowest?
+
+**iter 321 — 3,292 ms — at t ≈ 143.7 s** (see Plot 1, red star).
+
+It is +2.1 % over the median 3,224 ms. Looks like JIT warm-tail rather than a runtime stall.
+
+## Q3. Make sure it's not warm-up phase
+
+Confirmed it isn't.
+
+- Real denoising span: **t = 133 s → 793 s** (180 iters, blue cloud in Plot 1).
+- Pre-denoise (TextEncoder + VAE encode/decode): t = 0 → 133 s (orange band in Plot 1).
+- iter 321 sits ~10 s **after** real denoising begins → it is the first steady-state iter, hot-cache.
+- σ = 21.2 ms across 180 iters (0.66 % of median) — distribution is extremely tight (Plot 2). No straggler outliers.
+
+## What the plots show
+
+| File | What it shows |
+| --- | --- |
+| `1_iter_duration_over_time.png` | All 180 iters as scatter vs wall time. Warm-up phase shaded. Slowest iter starred. |
+| `2_iter_duration_histogram.png` | Steady-state iter distribution. σ = 21.2 ms, ~Gaussian. |
+| `3_stream7_one_iter.png` | Stream 7 timeline for the slowest iter, full + 200 ms zoom. **Blue (compute) and red (NCCL) strictly alternate on the same stream — no overlap possible.** |
+| `4_per_device_walltime.png` | Per-GPU wall-time stack: 47 % compute / 34 % NCCL / 18 % idle on every device. **Load is balanced — no single-GPU bottleneck.** |
+| `5_nccl_kernel_duration_hist.png` | NCCL kernel duration histogram. Two peaks at ~13 ms and ~33 ms → the two AllToAll4D message sizes (13 MiB pre-attn scatter, 40 MiB post-attn gather). |
+
+## Headline conclusion (already in report §1–§9)
+
+GPU 0..3 are all spending **34 % of wall time inside NCCL kernels** that are **serialized with compute on the same stream**. Compute/NCCL overlap is structurally 0 % (Plot 3 zoom). The 18 % idle is host-side launch + sync gap, not network. PCIe Gen4 × no NVLink × 4 ranks is the upstream cause; the two land-able levers are (a) TGATE (token-gate CFG-skip, landed at −22 %) and (b) CFG batching (predicted −8…−15 %).

profiles/perf.nsys-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46e24860eb6d98a9f504af3050401ab4f53632184827df27540232565f491bee
+size 633737449

profiles/perf.sqlite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b81a2b1282256ff5356903a8bb27a9c57ca21db81b52d6f841ea8883d6616a07
+size 2189783040

profiles/perf_compile.nsys-rep

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c012ba07b26054197dffeeeb2195f73765f65e58b8b6006cb5ea6f0eccecac2d
+size 530098795

profiles/perf_compile.sqlite

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec5a19dee22de14296aa6195e818d0e888df216d715ee7e2cd362489d4fa1f06
+size 1882554368